Agent construction and scheduling method supporting multiple simulation engines

Abstract

The application relates to the technical field of intelligent agent construction, and discloses an intelligent agent construction and scheduling method supporting multiple simulation engines, which comprises the following steps: acquiring a mainstream engine mode of an intelligent robot in a simulation process, and performing state optimization processing on the mainstream engine mode to output a simulation engine set; constructing a simulation intelligent agent supporting the running of the multiple simulation engines based on the simulation engine set, and running the simulation engine set to verify the behavior logic of the simulation intelligent agent, and optimizing the construction process of the simulation intelligent agent, stopping the optimization after the behavior logic meets the demand, and outputting an intelligent agent set of the intelligent robot containing several simulation intelligent agents; evaluating target fidelity required in the simulation process through real-time simulation task of the intelligent robot, scheduling the simulation intelligent agent from the intelligent agent set based on the target fidelity, and executing the simulation processing of the intelligent robot. The intelligent agent scheduling mechanism driven by the target fidelity ensures the convenient response of the simulation intelligent agent when facing different tasks.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent construction technology, and more specifically, to an intelligent agent construction and scheduling method that supports multiple simulation engines. Background Technology

[0002] Intelligent robots are autonomous devices with multiple functions such as perception, decision-making, learning, and execution. They can simulate human behavior and decision-making processes to a certain extent and perform tasks in complex environments. Intelligent robot simulation agents are intelligent agents constructed through computer simulation technology to simulate the behavior of intelligent robots in a virtual environment. They are usually used to test and optimize various situations and tasks that robots may encounter in real environments. The purpose is to predict and improve the performance of robots through simulation experiments in virtual environments, and reduce the risks and costs of conducting experiments in reality.

[0003] However, existing methods for constructing intelligent robot simulation agents typically do not fully utilize historical data to accurately identify key behavioral patterns, nor do they improve the adaptive performance of simulation agents through self-optimization techniques. Their multi-engine simulation and parallel processing capabilities are relatively limited, making it easy for simulation agents to encounter performance bottlenecks in complex tasks. Furthermore, existing simulation agents cannot achieve a precise target fidelity-driven scheduling mechanism in terms of multi-engine collaborative work and task scheduling, resulting in low resource scheduling efficiency and an inability to accurately match the capabilities and task requirements of simulation agents as needed.

[0004] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0005] To address the problems in related technologies, this invention proposes a method for constructing and scheduling intelligent agents that supports multiple simulation engines, thereby overcoming the aforementioned technical problems in existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows:

[0007] A method for constructing and scheduling intelligent agents that supports multiple simulation engines, including:

[0008] The matching degree between the simulation service requirements of intelligent robots and the simulation environment is quantified by using historical simulation process data, and the behavioral states used to describe the simulation process of intelligent robots are divided according to the quantification results.

[0009] Based on the behavioral state, the mainstream engine modes of the intelligent robot in the simulation process are obtained, and the state optimization processing of the mainstream engine modes is carried out through self-optimization operation technology to output a set of simulation engines.

[0010] Based on twin simulation and distributed theory, a simulation agent that supports the operation of multiple simulation engines is constructed based on a set of simulation engines, and the behavior logic of the simulation agent is verified by running the set of simulation engines.

[0011] The construction process of the simulated intelligent agent is optimized based on the behavioral logic verification results, and the optimization stops after the behavioral logic meets the requirements. The result is a set of intelligent agents for intelligent robots containing several simulated intelligent agents.

[0012] The required target fidelity during the simulation process is evaluated through real-time simulation tasks of intelligent robots. Based on the target fidelity, simulation agents are scheduled from the set of agents to perform intelligent robot simulation processing.

[0013] Preferably, the mainstream engine modes of the intelligent robot in the simulation process are obtained based on behavioral states, and the state optimization processing of the mainstream engine modes is performed through self-optimization operation technology, resulting in a set of output simulation engines including:

[0014] The behavior state is mined and processed to output a set of motifs that appear repeatedly and have information gain. The frequency of use, the transferability of the application process, and the business criticality are selected as the engine pattern screening conditions.

[0015] The attribute threshold points are used as the root nodes of the decision tree to divide the engine mode selection conditions into two subsets. The two subsets are then recursively processed to generate the engine selection decision tree for the intelligent robot during simulation.

[0016] Based on the cost complexity post-pruning technique, cross-validation is used to optimize the engine selection decision tree, and the optimized engine selection decision tree is used to select the mainstream engine mode from the modality set.

[0017] The mainstream engine modes are modeled, and the state space of the modeling results is mapped to the manifold space to adjust the engine modes, resulting in a set of simulation engines.

[0018] Preferably, the frequency used in the engine mode filtering criteria represents the frequency of occurrence of any simulation engine mode during the execution of intelligent robot simulation tasks, or the number of times or proportion of times it is used during the simulation process.

[0019] Application process portability refers to the efficiency and effectiveness of any simulation engine mode when it is migrated from one intelligent robot simulation task to another intelligent robot simulation task.

[0020] Business criticality indicates the importance or priority of any simulation engine mode when performing intelligent robot simulation tasks.

[0021] Preferably, the attribute threshold point is used as the root node of the decision tree to divide the engine mode selection conditions into two subsets, and the two subsets are recursively processed to generate the engine selection decision tree for the intelligent robot during simulation, including:

[0022] Based on usage frequency, application process portability, and business criticality, the engine pattern samples are sorted in ascending order of their attribute values, and the boundary points between adjacent heterogeneous engine pattern samples are found based on the sorting results.

[0023] Calculate the average class entropy of the boundary points, select the boundary point corresponding to the minimum average class entropy as the attribute threshold point, and use the attribute threshold point as the root node of the decision tree to divide the engine mode samples.

[0024] Based on the partitioning results, a sample subset is generated, and the optimal threshold for the engine mode filtering conditions in the sample subset is calculated for each subset. The sample subset is then recursively partitioned using the optimal threshold until the engine mode attributes are the same.

[0025] The decision tree is constructed based on a subset of samples, and the subtree sequence is generated by selecting nodes that meet the cost complexity function requirements by adjusting the pruning threshold variable. This results in the engine-selected decision tree.

[0026] Preferably, mainstream engine modes are modeled, and manifold learning and locally linear embedding techniques are combined to map the state space of the modeling results to the manifold space for engine mode adjustment, resulting in a set of simulation engines including:

[0027] Based on the state variables of various mainstream engine modes, the execution success rate and latency distribution are introduced to generate corresponding behavioral features. These behavioral features are then linked through a graph neural network to output the modeling results.

[0028] Based on the modeling results of each mainstream engine mode, analyze the start-stop and on / off states of the engine modes, and based on the analysis results, infer the synchronous alternation state of each mainstream engine mode when running simulation tasks.

[0029] Decision variables are determined based on synchronous alternating states, and objective functions are generated using decision variables and constraints to construct a multi-objective optimization model to optimize the state of the modeling results of mainstream engine modes.

[0030] By using the local linear embedding technique, a manifold mapping is performed on the high-dimensional state space of the optimized mainstream engine mode modeling results, and the state of the mainstream engine mode is adjusted to obtain a set of simulation engines.

[0031] Preferably, using local linear embedding technology, a manifold mapping is performed on the high-dimensional state space of the optimized mainstream engine mode modeling results, and the states of the mainstream engine modes are adjusted to obtain the simulation engine set including:

[0032] The high-dimensional state space of the optimized mainstream engine mode modeling results is analyzed, and the core feature vectors of the mainstream engine mode are extracted in the high-dimensional state space. The core feature vectors are then mapped to the low-dimensional manifold space by using the local linear embedding technique while preserving the inherent correlation of the mainstream engine modes.

[0033] In the low-dimensional popular space, the optimal solution of the mainstream engine mode is determined according to the distribution state of the core feature vector, and the mainstream engine mode is adjusted based on the optimal solution.

[0034] Based on the adjustment results, optimize the state of the mainstream engine modes in the low-dimensional manifold space that are far from the optimal solution, so as to ensure that the state of each mainstream engine mode tends to the optimal solution. The simulation engine set is obtained based on the adjustment results.

[0035] The beneficial effects of this invention are as follows:

[0036] 1. This invention establishes an adaptation relationship between behavioral states and simulation services through historical data-driven approaches, achieving fine alignment between simulation modeling and environmental dynamics. This enhances the targeted nature of agent construction. By combining mainstream engine pattern recognition and self-optimizing operation technologies, it ensures that each simulation agent has adaptable execution capabilities, effectively improving the adaptive performance of the simulation. Simultaneously, it utilizes twin simulation and distributed computing architecture to construct multi-engine simulation agents, enabling them to possess multi-dimensional parallel processing capabilities and the dynamic evolution capabilities of complex behaviors. This significantly enhances the realism and timeliness of the simulation process. Furthermore, through an agent scheduling mechanism driven by target fidelity, it achieves on-demand invocation, precision matching, and optimal performance dynamic scheduling, ensuring the ease of response of simulation agents when facing different tasks.

[0037] 2. By mining a set of motifs with information gain, this invention can identify key behavioral patterns that recur during the simulation process, providing high-quality support for subsequent engine pattern selection. Furthermore, by combining manifold learning and local linear embedding techniques to model and spatially adjust the engine patterns, it not only enhances the adaptability of the simulation engine patterns but also enables dynamic adjustment in complex environments, thereby improving the accuracy and efficiency of the entire simulation process.

[0038] 3. This invention integrates twin simulation technology with a set of simulation engines to build a scalable and parallel-running simulation agent architecture. This enables the simulation process to adapt to complex and ever-changing real-world task scenarios and forms a dynamic mapping between the virtual and real environments. Furthermore, the data acquired during operation can be used for behavioral logic verification through distributed theory, thereby achieving systematic evaluation and logical closed-loop verification of the simulation agent's performance and improving the simulation agent's task adaptability and real-time response capabilities. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a flowchart of an intelligent agent construction and scheduling method supporting multiple simulation engines according to an embodiment of the present invention. Detailed Implementation

[0041] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention.

[0042] According to embodiments of the present invention, a method for constructing and scheduling intelligent agents that supports multiple simulation engines is provided.

[0043] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the agent construction and scheduling method supporting multiple simulation engines according to an embodiment of the present invention includes:

[0044] Step S1: Quantify the matching degree between the simulation service requirements of the intelligent robot and the simulation environment using historical simulation process data, and classify the behavioral states used to describe the simulation process of the intelligent robot based on the quantification results.

[0045] In one embodiment, by collecting historical simulation data during the process of describing the behavioral state of an intelligent robot, information such as key performance indicators, sensor data, and execution trajectories of the robot performing tasks in different environments can be obtained. The service requirements (such as robot task objectives, performance requirements, and resource requirements) from the historical simulation process are compared with the simulation environment (such as the robot's physical environment, operating conditions, and interaction requirements) to quantify the degree of matching between the two. Specifically, feature vectors for requirements and the environment need to be defined. For example, requirements may include speed requirements, accuracy requirements, and task complexity, while the environment includes information such as obstacle distribution, ambient lighting, temperature, and humidity. The degree of matching between requirements and the environment is calculated using the feature vectors. If the characteristics of the requirements and the environment are highly consistent, it indicates that the robot can perform tasks well in that environment; otherwise, it indicates that there is a mismatch between the environment and the requirements, requiring adjustment or optimization.

[0046] Based on the quantification results of the matching between demand and environment, the behavioral states in the simulation process of intelligent robots are divided. The behavioral state refers to the different operating modes or states of the robot in the simulation process, such as standby state, task execution state, execution failure state, recovery state, etc. The decision-making strategy, task execution method and response to the environment adopted by the robot will be different in each state. In order to achieve quantitative division, it is necessary to perform clustering or classification analysis on historical simulation data, such as K-means clustering or decision tree methods. By analyzing the patterns of robot behavior in the data, different simulation processes are divided into several discrete state categories.

[0047] Step S2: Based on the behavioral state, obtain the mainstream engine modes of the intelligent robot in the simulation process, and perform state optimization processing on the mainstream engine modes through self-optimization operation technology to output the simulation engine set.

[0048] In one embodiment, the mainstream engine modes of the intelligent robot during the simulation process are obtained based on the behavioral state, and the state optimization processing of the mainstream engine modes is performed through self-optimizing operation technology, resulting in the output simulation engine set including:

[0049] The behavior state is mined and processed to output a set of motifs that appear repeatedly and have information gain. The frequency of use, the transferability of the application process, and the business criticality are selected as the engine pattern screening conditions.

[0050] The attribute threshold points are used as the root nodes of the decision tree to divide the engine mode selection conditions into two subsets. The two subsets are then recursively processed to generate the engine selection decision tree for the intelligent robot during simulation.

[0051] Based on the cost complexity post-pruning technique, cross-validation is used to optimize the engine selection decision tree, and the optimized engine selection decision tree is used to select the mainstream engine mode from the modality set.

[0052] The mainstream engine modes are modeled, and the state space of the modeling results is mapped to the manifold space to adjust the engine modes, resulting in a set of simulation engines.

[0053] In one embodiment, the attribute threshold point is used as the root node of the decision tree to divide the engine mode selection conditions into two subsets, and the two subsets are recursively processed to generate the engine selection decision tree for the intelligent robot during simulation, including:

[0054] Based on usage frequency, application process portability, and business criticality, the engine pattern samples are sorted in ascending order of their attribute values, and the boundary points between adjacent heterogeneous engine pattern samples are found based on the sorting results.

[0055] Calculate the average class entropy of the boundary points, select the boundary point corresponding to the minimum average class entropy as the attribute threshold point, and use the attribute threshold point as the root node of the decision tree to divide the engine mode samples.

[0056] Based on the partitioning results, a sample subset is generated, and the optimal threshold for the engine mode filtering conditions in the sample subset is calculated for each subset. The sample subset is then recursively partitioned using the optimal threshold until the engine mode attributes are the same.

[0057] The decision tree is constructed based on a subset of samples, and the subtree sequence is generated by selecting nodes that meet the cost complexity function requirements by adjusting the pruning threshold variable. This results in the engine-selected decision tree.

[0058] It should be explained that when acquiring mainstream engine modes, behavioral state data of intelligent robots in historical simulation tasks are mined to identify patterns that repeatedly appear in multiple simulation processes and have a significant impact on the task. These patterns are called motifs. These motifs have high information gain, meaning that they can significantly improve the accuracy or efficiency of simulation results during the simulation process. Behavioral states include path planning, different stages of task execution, and the interaction between the robot and the environment. Motifs refer to the key patterns that repeatedly appear in these behavioral states and affect the task results.

[0059] The specific selection criteria include three aspects: frequency of use, transferability of application process, and business criticality. Frequency of use measures the number of times a certain engine pattern appears in a task, reflecting its importance in actual simulation. Transferability of application process examines the effectiveness and efficiency of the pattern when it is transferred from one task to another. If an engine pattern performs well in multiple tasks, its transferability is high. Business criticality measures the key role that the engine pattern plays in completing the task, such as whether it directly affects the success or accuracy of the task.

[0060] Then, based on the screening criteria (usage frequency, application process transferability, business criticality), the sample engine patterns are sorted in ascending order of attribute values. The decision tree is recursively partitioned to generate an engine screening decision tree for optimization. When the engine screening decision tree is obtained, the generated decision tree is optimized. The goal of pruning is to remove the overfitted parts in the decision tree, reduce the complexity of the model, and improve its generalization ability. Cross-validation is used to evaluate the performance of different subtrees during the pruning process. By continuously adjusting the pruning threshold, the optimal decision tree model is finally selected.

[0061] In one embodiment, mainstream engine modes are modeled, and manifold learning and locally linear embedding techniques are combined to map the state space of the modeling results to the manifold space for engine mode adjustment, resulting in a set of simulation engines including:

[0062] Based on the state variables of various mainstream engine modes, the execution success rate and latency distribution are introduced to generate corresponding behavioral features. These behavioral features are then linked through a graph neural network to output the modeling results.

[0063] Based on the modeling results of each mainstream engine mode, analyze the start-stop and on / off states of the engine modes, and based on the analysis results, infer the synchronous alternation state of each mainstream engine mode when running simulation tasks.

[0064] Decision variables are determined based on synchronous alternating states, and objective functions are generated using decision variables and constraints to construct a multi-objective optimization model to optimize the state of the modeling results of mainstream engine modes.

[0065] By using the local linear embedding technique, a manifold mapping is performed on the high-dimensional state space of the optimized mainstream engine mode modeling results, and the state of the mainstream engine mode is adjusted to obtain a set of simulation engines.

[0066] In one embodiment, a manifold mapping is performed on the high-dimensional state space of the optimized mainstream engine mode modeling results using local linear embedding technology, and the states of the mainstream engine modes are adjusted to obtain a set of simulation engines including:

[0067] The high-dimensional state space of the optimized mainstream engine mode modeling results is analyzed, and the core feature vectors of the mainstream engine mode are extracted in the high-dimensional state space. The core feature vectors are then mapped to the low-dimensional manifold space by using the local linear embedding technique while preserving the inherent correlation of the mainstream engine modes.

[0068] In the low-dimensional popular space, the optimal solution of the mainstream engine mode is determined according to the distribution state of the core feature vector, and the mainstream engine mode is adjusted based on the optimal solution.

[0069] Based on the adjustment results, optimize the state of the mainstream engine modes in the low-dimensional manifold space that are far from the optimal solution, so as to ensure that the state of each mainstream engine mode tends to the optimal solution. The simulation engine set is obtained based on the adjustment results.

[0070] It should be explained that for the selected mainstream simulation engine modes, such as path planning, target tracking, and sensor fusion, key state variables during their operation are collected, including runtime input dimensions, CPU / GPU resource usage, average execution time (latency), and success rate (whether the target behavior is stably output). Based on this, the execution success rate and latency distribution are introduced as core parameters to measure the performance of behavioral characteristics. For example, the success rate of the path planning engine is 92%, and the mean of the latency distribution is 120ms and the variance is 25ms. 2Meanwhile, the above features are used as the behavior vectors of each engine mode. A relationship graph between engines is constructed, and the behavior vectors are structured and fused using a graph neural network to output the dynamic modeling results of each engine under the group relationship.

[0071] Based on the modeling results of GNN output, the start / stop status (whether it is active) and on / off status (whether it depends on other inputs / outputs) of each mainstream engine mode during simulation execution are analyzed. For example, the path planning engine has an activation rate of 80% throughout the simulation cycle, but the map update engine it depends on is only activated in the initial stage, resulting in a time on / off problem between the two. By analyzing whether each engine has synchronous or asynchronous alternating behavior with other engines in the simulation task, a synchronous alternating state matrix is ​​derived. This matrix quantifies the possible operational stage conflicts or cooperation relationships of each engine in the simulation process. For example, the path planning and obstacle avoidance engines are alternately activated in most task segments, while path planning and target recognition are usually synchronously activated.

[0072] Based on the aforementioned synchronous alternating states, a multi-objective optimization model is constructed, and several decision variables are extracted from it, such as whether to enable a certain engine, the maximum allowable latency range of the engine, and the resource limitations for parallel operation. At the same time, a set of constraints are set, such as resource consumption not exceeding 80%, success rate not less than 90%, and latency not exceeding 150ms. An optimization objective function is also set, such as a combination function that minimizes latency and maximizes success rate. Particle swarm optimization (PSO) or NSGA-II methods are used to optimize and search the state variables of each mainstream engine to obtain the optimized high-dimensional state space representation. For example, the optimized state of the path planning engine is (enabled = 1, latency = 110ms, success rate = 94%), and the state of the obstacle avoidance engine is (enabled = 1, latency = 95ms, success rate = 91%).

[0073] To reduce the computational cost of high-dimensional state modeling during scheduling, the Locally Linear Embedding (LLE) method is used to process the optimization results through manifold learning. While preserving the local neighborhood structure between the mainstream engine modes, LLE maps the high-dimensional state space to a low-dimensional manifold space (such as 2D or 3D), which facilitates subsequent state clustering and optimal state identification. In the low-dimensional space, the core feature vector distribution of each engine mode is identified, and the optimal state center is determined according to density clustering or maximum value analysis methods. For example, in the two-dimensional LLE embedding graph, the optimal state center of the path planning engine corresponds to coordinates (0.3, 0.7), while the current mapped state is (0.5, 0.9), indicating a deviation from the optimal state center.

[0074] Based on the optimal state center, the engine state is adjusted for all mapping results, so that each engine moves closer to its optimal state. During this process, if an engine is too far from the optimal region (e.g., exceeding a set threshold), δ =0.25), which will trigger an automatic parameter readjustment mechanism, such as shortening the engine's resource allocation window or reducing the execution frequency of non-critical behaviors, so that the state falls back to an acceptable range. Finally, through a series of mappings and adjustments, the resulting set of simulation engines can not only meet performance constraints, but also ensure the maximum efficiency of collaboration between engines.

[0075] Suppose that the current intelligent robot simulation task requires three engines: path planning (PP), obstacle avoidance (AO), and spatial mapping (SLAM). Historical data shows that their initial success rates are PP=91%, AO=88%, and SLAM=85%, with average latency of PP=130ms, AO=140ms, and SLAM=160ms, respectively. After graph neural network modeling, it is shown that PP and AO are highly coordinated, while SLAM and PP have some conflicts (overlapping in the update cycle). After synchronous alternation analysis and objective function optimization, the success rates of the three engines are improved to PP=94%, AO=91%, and SLAM=87%, respectively, and the latency is reduced to PP=110ms, AO=100ms, and SLAM=135ms, respectively. Furthermore, by using LLE to compress the three-dimensional state vector into a two-dimensional manifold space, the mapping results cluster around the optimal state solution. Through optimization strategies, a set of simulation engines that meets accuracy constraints and has optimal resource allocation is finally obtained.

[0076] Step S3: Based on twin simulation and distributed theory, construct a simulation agent that supports the operation of multiple simulation engines based on a set of simulation engines, and run the set of simulation engines to verify the behavioral logic of the simulation agent.

[0077] In one embodiment, based on twin simulation and distributed theory, a simulated intelligent agent supporting the operation of multiple simulation engines is constructed using a set of simulation engines as the foundation, and the behavioral logic of the simulated intelligent agent is verified by running the set of simulation engines, including:

[0078] Analyze the structural state of the intelligent robot to construct a digital 3D model, and verify the actual structural information of the intelligent robot within the digital 3D model to ensure that the virtual layout is consistent with the actual application and operation state;

[0079] Configure the parameter attributes and physical behavior of the intelligent robot for the verified digital 3D model, and use sensors and signals to realize automated control to ensure real-time synchronization of the virtual and real intelligent robot.

[0080] Based on the configured digital 3D model, twin simulation technology and a set of simulation engines are used to establish a mapping relationship between the virtual environment and the real environment, generating a simulated intelligent agent that supports the operation of multiple simulation engines;

[0081] The simulation tasks of the intelligent robot are randomly assigned to the simulated intelligent agent, and the process is executed by multiple simulation engines. The data obtained during the process is used to verify the behavioral logic of the simulated intelligent agent using distributed theory.

[0082] In one embodiment, based on the configured digital 3D model, a mapping relationship between the virtual environment and the real environment is established using twin simulation technology and a set of simulation engines to generate a simulation agent that supports the operation of multiple simulation engines, including:

[0083] Based on the configured digital 3D model, a motion control-driven design and establishment of an intelligent robot simulation digital workstation is carried out, and the configuration design of the running object is carried out according to the fully integrated automation platform.

[0084] The simulation integration test platform is generated by integrating the establishment and design results, and the kernel of the simulation integration test platform is optimized by calling script nodes using a programming method that combines engineering modeling and automated testing.

[0085] The kernel-optimized simulation integration test platform is combined with the digital 3D model and sent to the engineering modeling main panel, and the mapping relationship between the virtual environment and the real environment is established via Ethernet.

[0086] Based on virtual development technology using engineering modeling and parallel computing capabilities for automated testing, a simulation integration testing platform is configured under conditions of multiple simulation engines, resulting in a simulation intelligent agent that supports the operation of multiple simulation engines.

[0087] It should be explained that in the process of building a simulated intelligent agent, a complete digital 3D model of the robot is generated using tools such as computer-aided design (CAD). This model not only includes the robot's appearance, dimensions, and other geometric information, but also involves the connection relationships and working principles between various components, such as a six-degree-of-freedom robotic arm, transmission system, sensors, and actuators. The 3D model enables the simulation and verification of the robot's actual structure, ensuring that the digital model is consistent with the physical layout and behavior in the actual robot application scenario. The verified digital 3D model can lay the foundation for subsequent control and simulation work.

[0088] Simultaneously, various attributes are configured for the robot, such as speed, acceleration, mass, and coefficient of friction, defining the robot's physical behavior to ensure its performance in the simulation environment realistically reflects real-world conditions. Based on the configured digital 3D model, twin simulation technology and a set of simulation engines are used to establish a mapping relationship between the virtual and real environments, generating a simulated intelligent agent that supports the operation of multiple simulation engines. Twin simulation technology synchronizes the virtual and real worlds by creating a digital twin, establishing a dynamic mapping relationship between the virtual and real environments. By introducing multiple simulation engines, different physical phenomena or behaviors can be simulated, and the engines can be used to perform tasks. For the robot, the simulation engine in the virtual environment can simulate its path planning, action execution, collision detection, etc., while the simulation engine in the real environment maps these behaviors to the real world through hardware interfaces. This synchronization between virtual and reality provides reliable support for complex simulations, ensuring that the robot can smoothly transition between the two.

[0089] For example, when a robot performs an obstacle avoidance task or an object grasping task, the simulated intelligent agent, supported by multiple simulation engines, can simultaneously simulate multiple task scenarios, handle different physical interactions and task requirements. During task execution, the robot's motion trajectory and movement are adjusted in real time through sensors and signal feedback. The application of distributed theory aims to decompose the robot task into multiple sub-tasks and process them in parallel through multiple simulation engines. Parallel processing can greatly improve the efficiency of simulation tasks, while avoiding performance bottlenecks caused by excessive load on a single engine. The core idea of ​​distributed theory is to ensure that each engine can run independently when processing its own task through reasonable task allocation and parallel computing, and coordinate with other engines through synchronization mechanisms, thereby maintaining the stability and efficiency of the simulation process.

[0090] Based on the configured digital 3D model, a digital workstation for intelligent robot simulation is designed and built using motion control. This stage transforms the digital 3D model into a simulation workstation, which integrates a control system, sensor modules, and a computing platform. It can support the robot's full-range simulation in a digital environment. The motion control-driven design enables the robot to adjust its trajectory according to real-time input data during the simulation process, ensuring the smooth completion of the simulation task.

[0091] Step S4: Optimize the construction process of the simulated intelligent agent based on the behavioral logic verification results, and stop the optimization after the behavioral logic meets the requirements, and output a set of intelligent agents for intelligent robots containing several simulated intelligent agents.

[0092] In one embodiment, after the initially constructed simulated agent completes one or more rounds of simulation tasks, its behavioral logic execution results are verified. The verification dimensions mainly include core indicators such as task completion rate, behavioral stability, policy response latency, and behavioral chain integrity. Based on the deviation feedback in the behavioral logic verification, the agent construction parameters, such as task scheduling order, simulation engine priority, resource allocation strategy, perception update frequency, and action execution granularity, are automatically adjusted.

[0093] After optimization, the new simulated agent is rerun to perform the simulation task, and behavioral logic verification is executed again. This process constitutes a closed-loop mechanism for verification and optimization. In each iteration, the performance of the simulated agent in the task is recorded and compared with the preset behavioral logic threshold. If all logical requirements are met for N consecutive rounds (e.g., 3 rounds), or the average score is higher than the set upper limit (e.g., 95 points), it is considered to meet the task requirements, and the automatic stopping optimization mechanism is triggered to prevent overfitting or waste of resources.

[0094] Assume that three simulated intelligent agents A1, A2, and A3 are initially constructed to complete the task of avoiding random obstacles and transporting objects to a target point in a simulated factory environment. The behavioral logic settings include:

[0095] Obstacle avoidance response latency ≤200ms at any given time; total task completion time ≤12 seconds; operation sequence strictly follows perception, localization, obstacle avoidance, and transport; the entire process must not be interrupted or logically reversed. The system recorded the following verification results after the first simulation run:

[0096] A1: Completion time 10.5 seconds, obstacle avoidance response 180ms, complete behavioral logic;

[0097] A2: Completion time 13.2 seconds, obstacle avoidance response 240ms, one stage of behavioral logic missing;

[0098] A3: Completion time 11.8 seconds, obstacle avoidance response 310ms, path logic reversal exists;

[0099] Optimizations were made to A2 and A3. A2 increased the granularity of action execution and reduced the perception update cycle (from 200ms to 100ms). A3 increased the obstacle avoidance engine priority (from the default 0.5 to 0.8) and modified the path planning model to a lightweight path graph. After the second round of verification:

[0100] A2: Completion time 11.0 seconds, obstacle avoidance response 190ms, complete behavioral logic;

[0101] A3: Completion time 10.8 seconds, obstacle avoidance response 180ms, complete behavioral logic.

[0102] The records show that in two consecutive rounds, scores A1, A2, and A3 all achieved logical integrity and met performance requirements, with average scores of 97, 94, and 95 respectively (scoring dimensions including behavioral integrity 40%, timeliness 30%, and resource utilization efficiency 30%). This meets the conditions for terminating optimization, and A1, A2, and A3 are automatically included in the output set. The final output set of intelligent agents (A1, A2, and A3) can be deployed in different robot task environments, thus forming a closed-loop, controllable, and automatically iteratively optimized simulation intelligent agent construction mechanism, ensuring that the output intelligent agents have high robustness and logical reliability.

[0103] Step S5: Evaluate the target fidelity required during the simulation process through the real-time simulation task of the intelligent robot, and schedule the simulation intelligent agents from the set of intelligent agents based on the target fidelity to perform intelligent robot simulation processing.

[0104] In one embodiment, the required target fidelity during the simulation process is evaluated through a real-time simulation task of an intelligent robot, and simulation agents are scheduled from the agent set based on the target fidelity to perform intelligent robot simulation processing, including:

[0105] Acquire real-time simulation tasks of intelligent robots and analyze the dynamic changes in the environment, execution time limits, and resource consumption requirements during the operation of real-time simulation tasks as task requirement information;

[0106] The accuracy that the real-time simulation task needs to ensure during the simulation process is evaluated based on the task requirement information, which is used as the target fidelity, and simulation agents are scheduled from the agent set according to the target fidelity.

[0107] Once the required simulation agent is activated, it performs the corresponding intelligent robot simulation processing operations according to the real-time simulation task requirements and records the simulation processing process.

[0108] In one embodiment, evaluating the accuracy that a real-time simulation task needs to maintain during simulation based on task requirement information, using this as the target fidelity, and scheduling simulation agents from the agent set according to the target fidelity includes:

[0109] Predict the dynamic change trend of the real-time simulation task during the execution of the simulation based on the task requirement information, evaluate the accuracy requirement of the real-time simulation task based on the dynamic change trend, and convert the accuracy requirement into a target fidelity value.

[0110] The optimal accuracy improvement path is determined based on the target fidelity value using a fuzzy inference algorithm. A scheduling mechanism is then set up based on the accuracy improvement path to schedule simulated agents that meet the requirements from the agent set.

[0111] It should be explained that the goal of acquiring real-time simulation tasks of intelligent robots and analyzing the corresponding dynamic changes in the environment, execution time limits, and resource consumption requirements during the operation of real-time simulation tasks is to understand the operating environment of real-time simulation tasks and the resources required. Through sensor data, environmental monitoring information, and task requirement inputs (such as target localization, obstacle avoidance, grasping tasks, etc.), the dynamic characteristics of the task can be accurately identified. If the task requires the robot to move from a starting point to a target point, and there are dynamic obstacles along the way, then the time window of dynamic changes, the relative speed of obstacles, the robot's speed, and computing resources are analyzed to obtain detailed task requirement information. Task requirement information may include, but is not limited to: execution time limits during the simulation process (such as needing to complete the task within 15 seconds), resource consumption requirements (such as the computational load of each engine cannot exceed 70%), and accuracy requirements for real-time tasks (such as the error of path planning being less than 10cm).

[0112] For a given task, if the task requires the robot to locate and pick up an object, then the target fidelity will include positioning accuracy, grasping accuracy, and object tracking accuracy. The accuracy requirements of the task are transformed into target fidelity indicators, such as path planning accuracy needing to be kept within 5cm and target recognition accuracy needing to reach over 90%. Based on this target fidelity, simulated agents that can meet these accuracy requirements are selected from the optimized set of agents. If the task accuracy requirement is high, agents with higher resource consumption and stronger computing power are selected; if the accuracy requirement is moderate, agents with lighter computational load are selected.

[0113] Specific accuracy requirements are driven by the prediction of dynamic trends. By analyzing historical data and environmental information during the execution of real-time simulation tasks, the possible trends in the task can be predicted. For example, if the task involves dynamic obstacle avoidance, the robot will face obstacles with different speeds and directions. The trajectories and speeds of these obstacles are uncertain, so it is necessary to use historical data for trend prediction, such as the average speed and acceleration of the obstacles, to further calculate the accuracy requirements in the task.

[0114] The optimal accuracy improvement path is determined based on the target fidelity value using a fuzzy inference algorithm. The algorithm derives a reasonable accuracy improvement path based on the accuracy requirements of the target fidelity, such as path planning error and obstacle avoidance response time. For example, it infers that for a certain task, improving obstacle avoidance accuracy should prioritize increasing the sensor update frequency, while improving path planning accuracy requires increasing the scheduling of computing resources. Based on the analysis results, a scheduling mechanism is set up to schedule suitable simulated agents for task processing.

[0115] Once a suitable agent is selected from the agent set, the simulation agent will be launched according to the requirements of the real-time simulation task. At this time, the simulation agent will start to execute the task, such as path planning, obstacle avoidance, and grasping, and monitor the simulation process in real time, recording the agent's performance in executing the task, such as task completion time, resource consumption, and accuracy error.

[0116] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing and scheduling intelligent agents that supports multiple simulation engines, characterized in that, The method includes: The matching degree between the simulation service requirements of intelligent robots and the simulation environment is quantified by using historical simulation process data, and the behavioral states used to describe the simulation process of intelligent robots are divided according to the quantification results. The behavior state is mined and processed to output a set of motifs that appear repeatedly and have information gain. The frequency of use, the transferability of the application process, and the business criticality are selected as the engine pattern screening conditions. The attribute threshold points are used as the root nodes of the decision tree to divide the engine mode selection conditions into two subsets. The two subsets are then recursively processed to generate the engine selection decision tree for the intelligent robot during simulation. Based on the cost complexity post-pruning technique, cross-validation is used to optimize the engine selection decision tree, and the optimized engine selection decision tree is used to select the mainstream engine mode from the modality set. The mainstream engine patterns are modeled, and the state space of the modeling results is mapped to the manifold space to adjust the engine patterns, resulting in a set of simulation engines. The mainstream engine patterns refer to the patterns that repeatedly appear in multiple simulation processes and have an impact on the simulation task. Based on twin simulation and distributed theory, a simulation agent that supports the operation of multiple simulation engines is constructed based on a set of simulation engines, and the behavior logic of the simulation agent is verified by running the set of simulation engines. The construction process of the simulated intelligent agent is optimized based on the behavioral logic verification results, and the optimization stops after the behavioral logic meets the requirements. The result is a set of intelligent agents for intelligent robots containing several simulated intelligent agents. The required target fidelity during the simulation process is evaluated through real-time simulation tasks of intelligent robots. Based on the target fidelity, simulation agents are scheduled from the set of agents to perform intelligent robot simulation processing.

2. The method of claim 1, wherein, The frequency used in the engine mode filtering criteria represents the frequency of occurrence of any simulation engine mode during the execution of intelligent robot simulation tasks, or the number of times or proportion of times it is used during the simulation process. The application process portability refers to the efficiency and effectiveness of any simulation engine mode when it is migrated from one intelligent robot simulation task to another intelligent robot simulation task. The term "business criticality" refers to the importance or priority of any simulation engine mode when performing intelligent robot simulation tasks.

3. The method of claim 2, wherein, The process of using attribute threshold points as the root node of the decision tree to divide the engine mode selection criteria into two subsets, and then recursively processing these two subsets to generate the engine selection decision tree for the intelligent robot during simulation, includes: Based on usage frequency, application process portability, and business criticality, the engine pattern samples are sorted in ascending order of their attribute values, and the boundary points between adjacent heterogeneous engine pattern samples are found based on the sorting results. Calculate the average class entropy of the boundary points, select the boundary point corresponding to the minimum average class entropy as the attribute threshold point, and use the attribute threshold point as the root node of the decision tree to divide the engine mode samples. Based on the partitioning results, a sample subset is generated, and the optimal threshold for the engine mode filtering conditions in the sample subset is calculated for each subset. The sample subset is then recursively partitioned using the optimal threshold until the engine mode attributes are the same. The decision tree is constructed based on a subset of samples, and the subtree sequence is generated by selecting nodes that meet the cost complexity function requirements by adjusting the pruning threshold variable. This results in the engine-selected decision tree.

4. The method of claim 1, wherein, The process involves modeling mainstream engine modes, combining manifold learning and locally linear embedding techniques, and mapping the state space of the modeling results to the manifold space for engine mode adjustment. The resulting simulation engine set includes: Based on the state variables of various mainstream engine modes, the execution success rate and latency distribution are introduced to generate corresponding behavioral features. These behavioral features are then linked through a graph neural network to output the modeling results. Based on the modeling results of each mainstream engine mode, analyze the start-stop and on / off states of the engine modes, and based on the analysis results, infer the synchronous alternation state of each mainstream engine mode when running simulation tasks. Decision variables are determined based on synchronous alternating states, and objective functions are generated using decision variables and constraints to construct a multi-objective optimization model to optimize the state of the modeling results of mainstream engine modes. By using the local linear embedding technique, a manifold mapping is performed on the high-dimensional state space of the optimized mainstream engine mode modeling results, and the state of the mainstream engine mode is adjusted to obtain a set of simulation engines.

5. The method of claim 4, wherein, The process of using local linear embedding technology to perform manifold mapping on the high-dimensional state space of the optimized mainstream engine mode modeling results, and adjusting the states of the mainstream engine modes to obtain the simulation engine set includes: The high-dimensional state space of the optimized mainstream engine mode modeling results is analyzed, and the core feature vectors of the mainstream engine mode are extracted in the high-dimensional state space. The core feature vectors are then mapped to the low-dimensional manifold space by using the local linear embedding technique while preserving the inherent correlation of the mainstream engine modes. In the low-dimensional popular space, the optimal solution of the mainstream engine mode is determined according to the distribution state of the core feature vector, and the mainstream engine mode is adjusted based on the optimal solution. Based on the adjustment results, optimize the state of the mainstream engine modes in the low-dimensional manifold space that are far from the optimal solution, so as to ensure that the state of each mainstream engine mode tends to the optimal solution. The simulation engine set is obtained based on the adjustment results.

6. The method of claim 1, wherein, The method of constructing a simulated intelligent agent that supports the operation of multiple simulation engines based on a set of simulation engines, and verifying the behavioral logic of the simulated intelligent agent by running the set of simulation engines, includes: Analyze the structural state of the intelligent robot to construct a digital 3D model, and verify the actual structural information of the intelligent robot within the digital 3D model to ensure that the virtual layout is consistent with the actual application and operation state; Configure the parameter attributes and physical behavior of the intelligent robot for the verified digital 3D model, and use sensors and signals to realize automated control to ensure real-time synchronization of the virtual and real intelligent robot. Based on the configured digital 3D model, twin simulation technology and a set of simulation engines are used to establish a mapping relationship between the virtual environment and the real environment, generating a simulated intelligent agent that supports the operation of multiple simulation engines; The simulation tasks of the intelligent robot are randomly assigned to the simulated intelligent agent, and the process is executed by multiple simulation engines. The data obtained during the process is used to verify the behavioral logic of the simulated intelligent agent using distributed theory.

7. The method of claim 6, wherein, The configured digital 3D model utilizes twin simulation technology and a set of simulation engines to establish a mapping relationship between the virtual and real environments, generating a simulation agent that supports the operation of multiple simulation engines, including: Based on the configured digital 3D model, a motion control-driven design and establishment of an intelligent robot simulation digital workstation is carried out, and the configuration design of the running object is carried out according to the fully integrated automation platform. The simulation integration test platform is generated by integrating the establishment and design results, and the kernel of the simulation integration test platform is optimized by calling script nodes using a programming method that combines engineering modeling and automated testing. The kernel-optimized simulation integration test platform is combined with the digital 3D model and sent to the engineering modeling main panel, and the mapping relationship between the virtual environment and the real environment is established via Ethernet. Based on virtual development technology using engineering modeling and parallel computing capabilities for automated testing, a simulation integration testing platform is configured under conditions of multiple simulation engines, resulting in a simulation intelligent agent that supports the operation of multiple simulation engines.

8. The method for constructing and scheduling intelligent agents supporting multiple simulation engines according to claim 1, characterized in that, The step of evaluating the target fidelity required during the simulation process through a real-time simulation task using an intelligent robot, and scheduling simulation agents from the agent set based on the target fidelity to perform intelligent robot simulation processing includes: Acquire real-time simulation tasks of intelligent robots and analyze the dynamic changes in the environment, execution time limits, and resource consumption requirements during the operation of real-time simulation tasks as task requirement information; The accuracy that the real-time simulation task needs to ensure during the simulation process is evaluated based on the task requirement information, which is used as the target fidelity, and simulation agents are scheduled from the set of agents according to the target fidelity. Once the required simulation agent is activated, it performs the corresponding intelligent robot simulation processing operations according to the real-time simulation task requirements and records the simulation processing process.

9. The method for constructing and scheduling intelligent agents supporting multiple simulation engines according to claim 8, characterized in that, The process of evaluating the accuracy that the real-time simulation task needs to maintain during the simulation process based on task requirement information, as the target fidelity, and scheduling simulation agents from the agent set according to the target fidelity includes: Predict the dynamic change trend of the real-time simulation task during the execution of the simulation based on the task requirement information, evaluate the accuracy requirement of the real-time simulation task based on the dynamic change trend, and convert the accuracy requirement into a target fidelity value. The optimal accuracy improvement path is determined based on the target fidelity value using a fuzzy inference algorithm. A scheduling mechanism is then set up based on the accuracy improvement path to schedule simulated agents that meet the requirements from the agent set.

Images