Simulation implementation method, device and equipment for cloud-native intelligent agent action planning

Abstract

This application relates to a simulation implementation method, apparatus, and device for cloud-native intelligent agent action planning. The method includes: receiving action planning data of a target intelligent agent for a current task, the action planning data including at least one interactive action corresponding to the current task and an interaction object type corresponding to the at least one interactive action; a terminal determining candidate position information corresponding to an execution carrier based on the interaction object type corresponding to the at least one interactive action, the candidate position information representing at least one set of candidate positions corresponding to the at least one interactive action; and determining the interaction position corresponding to the at least one interactive action from the at least one set of candidate positions to obtain the target position information of the execution carrier. This method can efficiently and accurately identify and obtain all candidate positions corresponding to abstract interaction object types that can be interacted with by the execution carrier, comprehensively establishing a reliable mapping from abstract semantics to all potential interaction positions in a specific scenario.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a simulation implementation method, apparatus and equipment for cloud-native intelligent agent action planning. Background Technology

[0002] With the deep integration of cloud computing, artificial intelligence, and robotics, cloud-native agents are increasingly capable of handling complex, knowledge-intensive tasks. Their advancements in autonomous decision-making, learning, and human-computer interaction are bringing unprecedented application potential to fields such as operations and maintenance, customer service, and industrial automation. However, traditional physical testing often faces numerous challenges in verifying agent behavior logic and algorithm performance, including high data acquisition costs, difficulty in precisely controlling the environment, and limited security.

[0003] To effectively address the aforementioned challenges, building high-fidelity simulators has become a crucial step in the research and testing of intelligent agents. Simulators can cost-effectively generate large-scale, diverse scene data and provide a safe, controllable, and repeatable testing platform, thus supporting the initial verification of intelligent agents and their generated task program code. However, in the current simulation environment, accurately translating the abstract operational instructions output by the intelligent agent (such as semantic descriptions in the program code) into a sequence of physical actions achievable by the virtual execution vehicle, while ensuring that they semantically conform to expectations, remains a pressing technical challenge. Summary of the Invention

[0004] Therefore, it is necessary to provide a simulation implementation method, device, and equipment for cloud-native intelligent agent action planning that can improve the accuracy of simulation execution, addressing the aforementioned technical problems.

[0005] Firstly, this application provides a simulation implementation method for cloud-native intelligent agent action planning, including:

[0006] Receive action planning data of the target intelligent agent for the current task, the action planning data including at least one interactive action corresponding to the current task and the interactive object type corresponding to the at least one interactive action;

[0007] Based on the interaction object type corresponding to the at least one interactive action, candidate position information corresponding to the execution carrier is determined; the candidate position information is used to characterize at least one candidate position set corresponding to the at least one interactive action; the candidate position set includes candidate positions that the execution carrier is allowed to move to when performing the interactive action in the current simulation scene.

[0008] The interaction position corresponding to the at least one interaction action is determined from the at least one candidate position set to obtain the target position information of the execution carrier.

[0009] In one embodiment, determining the interaction position corresponding to the at least one interaction action from the at least one candidate position set to obtain the target position information of the execution carrier includes:

[0010] Obtain the initial interaction position of the execution carrier;

[0011] Based on the movement cost condition and the initial interaction position, the interaction position corresponding to the at least one interaction action is determined from the at least one candidate position set to obtain the target position information of the execution carrier; the movement cost condition is the condition that the total movement cost indicated by the target position information to be determined needs to satisfy; the total movement cost is determined based on the initial position of the execution carrier and the interaction position corresponding to the at least one interaction action.

[0012] In one embodiment, the mobility cost condition includes a condition that minimizes the total mobility cost indicated by the pending target location information;

[0013] The step of determining the interaction position corresponding to the at least one interaction action from the at least one candidate position set based on the movement cost condition and the initial interaction position, to obtain the target position information of the execution carrier, includes:

[0014] Based on the initial interaction position and the interaction position corresponding to the at least one interaction action, a cost objective function is constructed; the cost objective function is used to calculate the total movement cost.

[0015] With the goal of minimizing the cost objective function, the interaction position corresponding to the at least one interaction action is determined from the at least one candidate position set, thereby obtaining the target position information of the execution carrier.

[0016] In one embodiment, determining the interaction position corresponding to the at least one interaction action from the at least one candidate position set, with the goal of minimizing the cost objective function, to obtain the target position information of the execution carrier, includes:

[0017] According to the order of the interaction actions of the current task, traverse all candidate position combinations corresponding to each interaction action in turn;

[0018] During the process of traversing all candidate position combinations corresponding to each interaction action:

[0019] Calculate the current movement cost corresponding to the candidate positions selected before the current interaction action;

[0020] Calculate the minimum move cost corresponding to the unselected candidate position after the current interaction action;

[0021] If the sum of the current move cost and the minimum move cost is greater than the minimum of the current total move cost, the traversal continues from the next candidate position after the previous interaction.

[0022] In one embodiment, the total mobility cost includes at least one of the total path cost and the total penalty cost; the total penalty cost includes at least one of the total path coherence cost, the total semantic matching cost, and the total spatial and physical rationality cost.

[0023] The total path cost is used to characterize the path cost when the execution carrier moves between various interaction positions.

[0024] The total path coherence cost is used to characterize the irrationality of the connection of the execution carrier in the spatial path; the total semantic matching cost is used to characterize the degree of mismatch between the action semantics of the at least one interactive action and the selected interactive object instance; and the total spatial physical rationality cost is used to characterize the irrationality of the posture of the execution carrier and the physical environment when performing the action.

[0025] In one embodiment, determining the candidate location information corresponding to the execution carrier based on the interaction object type corresponding to the at least one interaction action includes:

[0026] From the current object instances in the current simulated scene, determine at least one set of candidate instances corresponding to the at least one interactive action; the set of candidate instances includes at least one candidate interactive object instance that matches the type of the interactive object.

[0027] By utilizing the positional matching relationship between the interaction object instance and the execution carrier, the interaction position of the execution carrier that matches the position of the candidate interaction object instance is determined, thereby obtaining the candidate position information corresponding to the execution carrier.

[0028] Secondly, this application also provides a simulation implementation device for cloud-native intelligent agent action planning, comprising:

[0029] A receiving module is used to receive action planning data of a target intelligent agent for a current task. The action planning data includes at least one interactive action corresponding to the current task and the type of interactive object corresponding to the at least one interactive action.

[0030] The determining module is used to determine candidate position information corresponding to the execution carrier based on the interaction object type corresponding to the at least one interaction action; the candidate position information is used to characterize at least one candidate position set corresponding to the at least one interaction action; the candidate position set includes candidate positions that the execution carrier is allowed to move to when performing the interaction action in the current simulated scene;

[0031] The target module is used to determine the interaction position corresponding to the at least one interaction action from the at least one candidate position set, and obtain the target position information of the execution carrier.

[0032] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the simulation implementation method for cloud-native intelligent agent action planning provided in the first aspect of this application.

[0033] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the simulation implementation method for cloud-native intelligent agent action planning provided in the first aspect of this application.

[0034] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the simulation implementation method for cloud-native intelligent agent action planning provided in the first aspect of this application.

[0035] The aforementioned simulation implementation method, apparatus, and device for cloud-native intelligent agent action planning involves a terminal receiving action planning data of a target intelligent agent for a current task. This action planning data includes at least one interactive action corresponding to the current task and the interaction object type corresponding to the at least one interactive action. Based on the interaction object type corresponding to the at least one interactive action, the terminal determines candidate position information corresponding to the execution carrier. The candidate position information is used to characterize at least one set of candidate positions corresponding to at least one interactive action, and the terminal determines the interaction position corresponding to at least one interactive action from the at least one set of candidate positions to obtain the target position information of the execution carrier. This embodiment of the application, by receiving and parsing the interaction object type in the intelligent agent planning data, efficiently and accurately identifies and obtains all candidate positions corresponding to the abstract interaction object type that can be interacted with by the execution carrier. It can locate all candidate position information matching the semantic type in the simulated scene, comprehensively establishing a reliable mapping from abstract semantics to all potential interaction positions in the specific scene, eliminating the semantic ambiguity caused by object polysemy, and ensuring that the starting point for subsequent action execution is accurate. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1This is an application environment diagram of a simulation implementation method for cloud-native intelligent agent action planning in one embodiment;

[0038] Figure 2 This is a flowchart illustrating a simulation implementation method for cloud-native intelligent agent action planning in one embodiment.

[0039] Figure 3 This is a schematic diagram of a simulator performing a visual animation based on a program in one embodiment;

[0040] Figure 4 This is a schematic diagram of the process for obtaining candidate location information in one embodiment;

[0041] Figure 5 This is a schematic diagram of the process for obtaining target location information in one embodiment;

[0042] Figure 6 This is a schematic diagram of the process for obtaining target location information in another embodiment;

[0043] Figure 7 This is a flowchart illustrating the simulation implementation method of cloud-native intelligent agent action planning in another embodiment;

[0044] Figure 8 This is a structural block diagram of a simulation implementation device for cloud-native intelligent agent action planning in one embodiment.

[0045] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0047] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0048] The simulation implementation method for cloud-native intelligent agent action planning provided in this application embodiment can be applied to, for example... Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on the cloud or other network servers. Terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can include smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, head-mounted devices, etc. A simulator can be deployed in terminal 102, which can create and maintain a virtual, interactive environment that can simulate real-world physical laws (such as gravity, collisions, and friction) and allow virtual entities (such as intelligent agents, robots, and objects) to perceive, move, and interact within it. Server 104 can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services. Server 104 can be the backend server for the aforementioned simulator, providing backend services to the simulator.

[0049] The simulation implementation method for cloud-native intelligent agent action planning provided in this application embodiment can be executed by computer equipment, which refers to electronic equipment with data computing, processing, and storage capabilities. Figure 1 Taking the application environment shown as an example, the simulation implementation method of cloud-native intelligent agent action planning can be executed by terminal 102 alone (such as the simulator installed and running in terminal 102 executing the simulation implementation method of cloud-native intelligent agent action planning), or the simulation implementation method of cloud-native intelligent agent action planning can be executed by server 104 alone, or terminal 102 and server 104 can interact and cooperate to execute it. This application does not limit this.

[0050] In one exemplary embodiment, such as Figure 2 As shown, a simulation implementation method for cloud-native intelligent agent action planning is provided, which is then applied to... Figure 1 Taking the terminal in the example, the explanation includes the following steps 202 to 206. Wherein:

[0051] Step 202: Receive action planning data of the target intelligent agent for the current task.

[0052] The action planning data includes at least one interactive action corresponding to the current task and the type of interactive object corresponding to at least one interactive action.

[0053] In this embodiment, target agents can generate action planning data in response to current task instructions. This action planning data can be used to characterize the abstract high-level instructions output by the target agents based on their understanding and decision-making capabilities regarding the task, aiming to guide the execution vehicle to complete the current task in a simulated scenario. The interactive actions in the action planning data refer to the discrete and continuous behaviors that the target agents expect the execution vehicle to complete in the simulated environment, and the interaction object type refers to the abstract category or semantic label of the object involved in the interactive action.

[0054] For example, please refer to Figure 3 Taking action-oriented cloud-native agents as an example, the agents receive action-related task instructions, which include an action name (watching TV) and an action description (taking the juice from the dining table to the coffee table in the living room and relaxing by watching TV on the sofa in the living room). The agents generate action planning data based on the received task instructions. The terminal's simulator receives this action planning data from the agents. In this action planning data, the first interaction action could be walking towards the dining table, and the interaction object type could be the dining table; the second interaction action could be grabbing the juice, and the interaction object type could be the juice; the third interaction action could be walking to the living room, and the interaction object type could be the living room.

[0055] Step 204: Based on the interaction object type corresponding to at least one interaction action, determine the candidate position information corresponding to the execution carrier.

[0056] The candidate location information is used to characterize at least one set of candidate locations corresponding to at least one interactive action. The set of candidate locations includes interactive locations that the execution vehicle is allowed to move to when performing an interactive action in the current simulated scene.

[0057] The candidate location information in this application embodiment can be used to indicate all potential spatial coordinates, regions, or posture sets generated by the terminal in the current simulated scene after identification and positioning for the received interaction object type, which are available for the execution carrier to perform corresponding interactive actions. Specifically, each interactive action can correspond to a candidate location set, which may include at least one candidate location. This candidate location represents a specific interactive location that is allowed to be moved to for performing the corresponding interactive action and possesses physical reachability and interaction conditions.

[0058] For example, the terminal determines at least one candidate position of the execution carrier for each interactive action and its corresponding interactive object type, and obtains a set of candidate positions corresponding to the interactive action, ensuring a comprehensive and accurate mapping from abstract semantic instructions to all specific, interactive physical positions.

[0059] In one possible implementation, for example, for the first interactive action "walk towards" and the interactive object type "table", the terminal can calculate all feasible locations near the table that the execution vehicle can "walk towards" and reach without collision, forming a candidate location set corresponding to the table. For the second interactive action "grab" and the interactive object type "juice", the terminal can determine a candidate location set corresponding to "juice" that allows the "grab" action to be performed based on the arm kinematic model of the execution vehicle.

[0060] Step 206: Determine the interaction position corresponding to at least one interaction action from at least one candidate position set to obtain the target position information of the execution carrier.

[0061] The target location information can represent the target interaction position corresponding to each interactive action performed by the execution vehicle, and it is used to indicate the action execution of the execution vehicle in the simulated scene.

[0062] In the aforementioned simulation implementation method for cloud-native intelligent agent action planning, the terminal receives action planning data of the target intelligent agent for the current task. This action planning data includes at least one interactive action corresponding to the current task and the interaction object type corresponding to the at least one interactive action. Based on the interaction object type corresponding to the at least one interactive action, the terminal determines the candidate position information corresponding to the execution carrier. The candidate position information is used to characterize at least one candidate position set corresponding to at least one interactive action, and the terminal determines the interaction position corresponding to at least one interactive action from the at least one candidate position set to obtain the target position information of the execution carrier. This embodiment of the application, by receiving and parsing the interaction object type in the intelligent agent planning data, efficiently and accurately identifies and obtains all candidate positions corresponding to the abstract interaction object type that can be interacted with by the execution carrier. It can locate all candidate position information matching the semantic type in the simulated scene, comprehensively establishing a reliable mapping from abstract semantics to all potential interaction positions in the specific scene, eliminating the semantic ambiguity problem caused by object polysemy, and ensuring that the starting point of subsequent action execution is accurate.

[0063] In practical applications, step 204 is not implemented in only one way. In one exemplary embodiment, such as... Figure 4 As shown, step 204 includes steps 402 to 404. Wherein:

[0064] Step 402: Determine at least one set of candidate instances corresponding to at least one interactive action from the current object instances in the current simulated scene.

[0065] The candidate instance set includes at least one candidate interaction object instance that matches the interaction object type. The candidate instance set refers to the set of all concrete physical object instances that semantically match the received interaction object type, identified, filtered, and collected in the current simulation scene.

[0066] For example, when the interaction action is "walk towards" and the interaction object type is "dining table", the terminal can traverse all defined entity objects in the internal data structure or object list of the current simulation scene, and filter out all object instances marked as "dining table" by comparing the semantic label, category ID or preset attributes of each entity object (e.g., "dining table A" in the living room and "dining table B" in the kitchen) to obtain a set of candidate instances for the interaction action "walk towards the dining table".

[0067] Step 404: Using the position matching relationship between the interaction object instance and the execution carrier, determine the interaction position of the execution carrier that matches the position of the candidate interaction object instance, and obtain the candidate position information corresponding to the execution carrier.

[0068] Among them, the position matching relationship refers to the rules or models that are pre-configured or calculated, which indicate the spatial position, posture or relative geometric conditions that the execution vehicle or its specific interactive components (such as end effectors, body center of gravity, etc.) must satisfy relative to the interactive object instance when successfully performing a specific interactive action. The position matching relationship includes the mapping relationship between the position of the interactive object instance and the interactive position of the execution vehicle.

[0069] For example, for each set of candidate instances corresponding to an interactive action, the terminal can calculate the interactive position that the execution carrier can reach and that satisfies the interactive conditions based on a preset position matching relationship and the position of each candidate interactive object instance in the set of candidate instances, thus obtaining the set of candidate positions corresponding to each interactive action.

[0070] For example, the terminal can search for all identified "table" entities in the current simulated scene and calculate all feasible positions that the execution vehicle can "walk towards" and reach, located near each table entity and without collision, forming a candidate position set for the interactive action "walk towards the table". For the second interactive action "grab" and the interactive object type "juice", the terminal can identify all "juice" entities and, based on the arm kinematic model of the execution vehicle, determine the position corresponding to each "juice" entity that allows the "grab" action to be performed, obtaining a candidate position set for the interactive action "grab juice".

[0071] In this embodiment, by utilizing the position matching relationship between the interactive object instance and the execution carrier, the position of the execution carrier that matches the position of the candidate interactive object instance can be determined, which can efficiently and accurately integrate all possible and selectable candidate execution carrier position information.

[0072] In some embodiments, for each interactive action and its corresponding set of candidate locations, the terminal can determine the final target location based on preset rules or conditions, such as combining contextual information, cost functions (e.g., closest distance, minimum path cost, or compliance with specific semantic constraints), to ensure that each interactive action has a unique and clear physical execution target, thereby guaranteeing the semantic correctness and spatial coherence of subsequent simulation actions.

[0073] In one exemplary embodiment, such as Figure 5 As shown, step 206 includes steps 502 to 504. Wherein:

[0074] Step 502: Obtain the initial interaction position of the execution carrier.

[0075] The initial interaction position refers to the physical position or posture of the execution vehicle before it begins executing the first interaction action. This position is the starting point of the execution vehicle in the current simulation scene and serves as the benchmark for subsequently calculating the movement cost from the current position to the first target interaction position.

[0076] Step 504: Based on the movement cost condition and the initial interaction position, determine the interaction position corresponding to at least one interaction action from at least one candidate position set to obtain the target position information of the execution carrier.

[0077] The movement cost condition is the condition that the total movement cost indicated by the target location information to be determined must meet. For example, it is a preset constraint or optimization objective that the total movement cost indicated by the target location information to be determined must meet. The movement cost condition may include minimizing the total movement cost, or the total movement cost being lower than a preset cost threshold.

[0078] The total movement cost is determined based on the initial interaction position of the execution vehicle and the interaction position corresponding to at least one interaction action. It may include the movement cost of the execution vehicle from the initial interaction position to the first interaction position and the movement cost between each successive interaction position.

[0079] For example, for each interactive action, the terminal can select a candidate location from the candidate location set corresponding to the interactive action as the interaction location for the execution carrier to perform the interactive action, thereby obtaining an interactive location sequence. If the total movement cost indicated by the interactive location sequence satisfies the movement cost condition, the interactive location sequence is determined as the target location sequence, that is, the target location corresponding to each interactive action, thus obtaining the target location information of the execution carrier. The movement cost condition can be used to indicate that the total movement cost is less than a preset cost threshold, or it can also be used to indicate the interactive location sequence with the minimum total movement cost.

[0080] In this embodiment, by using the movement cost condition, a candidate position can be determined from the candidate position set corresponding to each interactive action as the interaction position of that interactive action, so that the movement cost of the execution carrier between each interaction position meets the preset condition, effectively solving the ambiguity of the agent's abstract instructions in multi-instance scenarios, and ensuring that the finally determined target position is optimal or meets specific constraints.

[0081] In one exemplary embodiment, such as Figure 6 As shown, step 504 includes steps 602 to 604. Wherein:

[0082] Step 602: Construct a cost objective function based on the initial interaction position and the interaction position corresponding to at least one interaction action.

[0083] The cost objective function can be used to calculate the total cost of movement.

[0084] In one possible implementation, the total mobility cost may include at least one of the total path cost and the total penalty cost; the total penalty cost may include at least one of the total path coherence cost, the total semantic matching cost, and the total spatial-physical rationality cost.

[0085] The total path cost is used to characterize the path cost when the execution vehicle moves between various interaction locations.

[0086] Among them, the total path coherence cost is used to characterize the irrationality of the connection of the execution carrier on the spatial path; the total semantic matching cost is used to characterize the degree of mismatch between the action semantics of at least one interactive action and the selected interactive object instance; and the total spatial physical rationality cost is used to characterize the irrationality of the posture of the execution carrier and the physical environment when performing the action.

[0087] For example, the terminal can use the initial interaction position of the execution carrier and the interaction positions corresponding to each interaction action as input variables of the cost objective function to calculate the total movement cost, which may include path movement cost and reasonableness penalty for each interaction step. This cost objective function will serve as the standard for optimal decision-making and path selection in subsequent steps.

[0088] Step 604: With the goal of minimizing the cost objective function, determine the interaction position corresponding to at least one interaction action from at least one candidate position set to obtain the target position information of the execution carrier.

[0089] For example, the terminal can select a candidate position from the candidate position set corresponding to each interactive action as the initial undetermined position corresponding to each interactive action; based on the candidate position set corresponding to each interactive action, the terminal updates the initial undetermined position of each interactive action using each candidate position in the candidate position set to obtain the target position corresponding to each interactive action, that is, determine the target position information so as to minimize the total movement cost indicated by the target position information.

[0090] In practical applications, there is no single way to implement step 604. The terminal can use a search algorithm to perform calculations in a state space consisting of a set of candidate locations. This involves calculating the cost objective function value for different paths or sequences and minimizing that value as the optimization objective. During the search process, pruning strategies may be incorporated. For example, if the total movement cost of the current path is already higher than the known optimal solution, the exploration of that path can be abandoned to improve search efficiency.

[0091] The following section will further explain this application in conjunction with the cost objective function and the search algorithm.

[0092] In one exemplary embodiment, such as Figure 7 As shown, a simulation implementation method for cloud-native intelligent agent action planning is provided, the method including:

[0093] Step A1: The terminal determines the candidate instance set for each interactive action based on the action planning data of the target intelligent agent, and determines the candidate interaction position of the execution carrier corresponding to each candidate object instance in the candidate instance set.

[0094] Suppose a program has a total of Each step describes an abstract operation, such as "walking to the TV" or "turning on the TV".

[0095] For steps :

[0096] ;

[0097] in, Represents the set of candidate instances. Indicates the current step The kth candidate object instance, such as which specific TV or sofa in the virtual scene.

[0098] At the same time, define The execution vehicle of the Agents and the current step The interaction position that should be in when interacting with the kth candidate object instance.

[0099] For each step, a suitable candidate object instance is found. The overall goal is to find a sequence of candidate instances that makes the entire animation execution not only semantically reasonable but also spatially coherent and natural. The sequence of candidate instances can be represented as: .

[0100] Step A2: Based on the initial interaction position and the interaction position corresponding to at least one interaction action, construct the cost objective function.

[0101] To make the optimal choice among multiple candidate solutions, a global cost function is constructed, the core of which is the movement cost of agents between consecutive actions. Specifically, the cost objective function can be constructed as follows:

[0102] ;

[0103] in, This represents the path cost from the initial interaction position to the interaction position of the first interaction action. Indicates from the first Step interaction position to the first Path cost between interaction locations Represents the total movement cost function. It can represent the total penalty cost function, which can be a penalty function for each mapping candidate, used to measure whether the selected interaction position, etc., is reasonable (e.g., some positions appear unnatural or dangerous in actual operation).

[0104] To ensure that reasonable object interaction positions are selected when planning actions and executing animations in the simulator, a total penalty cost function is used. Penalizes unnatural animations. The function uses continuous values... This indicates the degree of irrationality; a larger value indicates a more irrational choice made in the current step. The total penalty cost function can be expressed by the following formula:

[0105] ;

[0106] The total penalty cost function includes the total path coherence cost function, which is used to calculate the irrationality of the connection of the execution carrier on the spatial path. The total path coherence cost value of the total path coherence cost function includes the path coherence cost value between each continuous interaction position, which can be determined based on the candidate interaction position selected by the previous interaction action, the candidate object instance corresponding to the candidate interaction position of the previous interaction action, the candidate interaction position selected by the current interaction action, and the candidate object instance corresponding to the candidate interaction position of the current interaction action.

[0107] For example, The penalty represents the continuity of the path, characterizing the logical connection between the current action and the previous action in the spatial path. If the current position cannot be reached from the previous position (e.g., due to a wall blocking the way or being outside the reachable range), the path is considered discontinuous and the highest penalty is applied; if the path is feasible, a continuous penalty value from 0 to 1 is applied based on the distance, with higher penalties for greater distances.

[0108] The total penalty cost function includes a total semantic matching cost function, which is used to calculate the degree of mismatch between the action semantics of at least one interactive action and the selected interactive object instance. The total semantic matching cost value of the total semantic matching cost function may include the semantic matching cost value corresponding to each interactive position, which can be determined based on the action semantics of each interactive action and the candidate object instance corresponding to the candidate interactive position selected by each interactive action.

[0109] For example, This represents a semantic matching penalty, measuring the degree of match between the action's semantics and the selected interaction object. If the action... semantic requirements and object instances If it doesn't match, it's considered unnatural. For example, the action "sit" usually only applies to objects that can be sat on, such as chairs and sofas, and should not be applied to objects that do not have this function, such as "television".

[0110] The total penalty cost function includes the total spatial-physical reasonable cost function, which is used to calculate the irrationality of the posture and physical environment of the execution vehicle when performing the action. The total spatial-physical reasonable cost value of the total spatial-physical reasonable cost function can include the spatial-physical reasonable cost value corresponding to each interaction position, which can be determined according to the action semantics of each interaction action, the candidate interaction position selected by each interaction action, and the candidate object instance corresponding to the candidate interaction position selected by each interaction action.

[0111] For example, Represents a spatial physical rationality penalty, measuring the agent's position. Execute action With object Does the agent's posture conform to physical and environmental constraints? This includes whether the agent is standing on a legal surface (such as the ground or a platform) rather than suspended in the air, whether it is colliding with or intersecting with furniture or walls, and whether its movements and postures are within the animation constraints.

[0112] Indicates the first The atomic action type of a step (e.g., "Sit" means to sit down, "walk" means to walk).

[0113] , , This represents the weight coefficients in the penalty function. Initially, a heuristic setting is used, with manual verification of its validity in conjunction with animation execution. .

[0114] Specifically, the cost of the shortest path It can be represented as:

[0115] ;

[0116] in, From point Time The set of paths that traverse the edges. It is the weight of each edge (often Euclidean distance).

[0117] By minimizing It can select a set of mapping sequences that ensures both low overall movement path cost and reasonable interaction at each step. The key to path disambiguation is to select the scheme that optimizes continuous movement (i.e., movement cost) when faced with multiple candidate mappings, thereby avoiding contradictory or incoherent walking routes in the same task.

[0118] Step A3: Solve the above cost objective function using search and backtracking methods to obtain the target position corresponding to each interaction action.

[0119] The terminal can sequentially traverse all candidate position combinations corresponding to each interaction action according to the order of the interaction actions of the current task. During the process of traversing all candidate position combinations corresponding to each interaction action: calculate the current movement cost value corresponding to the candidate position selected before the current interaction action; calculate the minimum movement cost value corresponding to the candidate position not selected after the current interaction action; if the sum of the current movement cost value and the minimum movement cost value is greater than the minimum value of the current total movement cost, continue traversing from the next candidate position of the previous interaction action.

[0120] For example, by constructing a search tree, expanding all candidate combinations at each step sequentially from the root, and backtracking when it is detected that the current path can no longer form a feasible solution (e.g., it will inevitably lead to high cost or conflict), the simulator can find an executable and coherent complete sequence in a short time.

[0121] First, define the state function, and let the state... Indicates the preceding The selected candidate solutions.

[0122] For a state A partial cost function can be defined:

[0123] ;

[0124] Furthermore, its lower bound function is:

[0125] ;

[0126] Wherein, the lower bound function Used to estimate the minimum additional cost required to complete all steps from the current state (the sum of the minimum possible cost and minimum penalty for each of all undetermined steps).

[0127] Let the cost of the current optimal solution be... Define a recursive function:

[0128] ;

[0129] The two conditions for this recursive function are: Condition 1, when... The current state This is a complete candidate solution; terminate the operation; Condition 2, if Then update and save. This is the current optimal solution.

[0130] For the Each candidate in the step Construct extended states:

[0131] ;

[0132] The pruning conditions are: if Then recursive call Otherwise, prune the branch (i.e., backtrack). If the total cost of the branch cannot be better than the currently known optimal solution, even if the optimal expansion is chosen later, then skip the search for that branch.

[0133] Using the above methods, a backtracking algorithm can be used to quickly find a solution that minimizes the global cost in the entire candidate space, thereby realizing the mapping from abstract operations to actual animation actions, eliminating ambiguity on the path, and ensuring the continuity between consecutive actions.

[0134] Step A4: Perform a visual assessment based on the determined target location information.

[0135] From program to actual animation action, execute all necessary interactive operations, evaluate whether the animation based on agents and program can fulfill the task requirements, that is, judge the correctness and executability of the program through animation, and complete the visual evaluation.

[0136] LCS (Longest Common Subsequence) is used to evaluate the correctness and short-term order consistency of action generation. It measures the length of the longest common subsequence between the generated action sequence and the "ground truth" action sequence, representing the proportion of the longest common subsequence to the ground truth action sequence. Its calculation method is as follows: .

[0137] EG Correctness (Environment Graph Correctness) measures whether the environment graph after executing the generated action plan is consistent with the environment state after executing the ground truth action plan. Figure 1 Therefore, this involves semantically evaluating whether the action plan correctly changes the environmental state. The calculation method is as follows:

[0138] .

[0139] For the post-execution environment state graph, extract the changed node set and edge set, including generated and ground truth values. Calculate the intersection-union ratio (IUU) of the nodes and the changed set respectively. .

[0140] In this embodiment, by constructing a cost function-based mapping scheme and a search and backtracking algorithm, automatic object mapping and path disambiguation are achieved from the program to the animation execution process, ensuring the spatial coherence and rationality between continuous actions, thereby improving the accuracy and practicality of agent action planning in the simulator and ensuring the reliability and effectiveness of the evaluation results.

[0141] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0142] Based on the same inventive concept, this application also provides a simulation implementation device for cloud-native intelligent agent action planning, which implements the simulation implementation method for cloud-native intelligent agent action planning described above. The solution provided by this device is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more embodiments of the simulation implementation device for cloud-native intelligent agent action planning provided below can be found in the limitations of the simulation implementation method for cloud-native intelligent agent action planning described above, and will not be repeated here.

[0143] In one exemplary embodiment, such as Figure 8 As shown, a simulation implementation device for cloud-native intelligent agent action planning is provided, including: a receiving module 802, a determining module 804, and a target module 806, wherein:

[0144] The receiving module 802 is used to receive action planning data of the target intelligent agent for the current task. The action planning data includes at least one interactive action corresponding to the current task and the type of interactive object corresponding to the at least one interactive action.

[0145] The determining module 804 is used to determine candidate position information corresponding to the execution carrier based on the interaction object type corresponding to the at least one interactive action; the candidate position information is used to characterize at least one candidate position set corresponding to the at least one interactive action; the candidate position set includes candidate positions that the execution carrier is allowed to move to when performing the interactive action in the current simulated scene.

[0146] The target module 806 is used to determine the interaction position corresponding to the at least one interaction action from the at least one candidate position set, and obtain the target position information of the execution carrier.

[0147] The modules in the aforementioned simulation implementation device for cloud-native intelligent agent action planning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0148] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and databases. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media to run. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a simulation method for cloud-native intelligent agent action planning.

[0149] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0150] In one exemplary embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the simulation implementation method for cloud-native intelligent agent action planning provided in the embodiments of this application.

[0151] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of the simulation implementation method for cloud-native intelligent agent action planning provided in the embodiments of this application are implemented.

[0152] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the simulation implementation method for cloud-native intelligent agent action planning provided in the embodiments of this application.

[0153] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0154] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0155] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0156] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A simulation implementation method for cloud-native intelligent agent action planning, characterized in that, The method includes: Receive action planning data of the target intelligent agent for the current task, the action planning data including at least one interactive action corresponding to the current task and the interactive object type corresponding to the at least one interactive action; Based on the interaction object type corresponding to the at least one interactive action, candidate position information corresponding to the execution carrier is determined; the candidate position information is used to characterize at least one candidate position set corresponding to the at least one interactive action; the candidate position set includes candidate interactive positions that the execution carrier is allowed to move to when performing an interactive action in the current simulation scene. The initial interaction position of the execution vehicle is obtained. Based on the movement cost condition and the initial interaction position, the interaction position corresponding to the at least one interaction action is determined from the at least one candidate position set to obtain the target position information of the execution vehicle. The movement cost condition is the condition that the total movement cost indicated by the target position information to be determined must satisfy. The total movement cost includes the total path cost and the total penalty cost. The total path cost characterizes the path cost when the execution carrier moves between various interaction locations. The total penalty cost includes the total path coherence cost, the total semantic matching cost, and the total spatial and physical rationality cost. The method further includes: Based on the candidate interaction position selected in the previous interaction, the candidate interaction object instance corresponding to the candidate interaction position in the previous interaction, the candidate interaction position selected in the current interaction, and the candidate interaction object instance corresponding to the candidate interaction position in the current interaction, the total path coherence value is determined; the total path coherence value is used to characterize the irrationality of the connection of the execution carrier in the spatial path. The total semantic matching cost is determined based on the action semantics of each interaction action and the candidate interaction object instance corresponding to the candidate interaction position selected by each interaction action; the total semantic matching cost is used to characterize the degree of mismatch between the action semantics of each interaction action and the selected candidate interaction object instance. Based on the action semantics of each interactive action, the candidate interaction positions selected by each interactive action, and the candidate interaction object instances corresponding to the candidate interaction positions selected by each interactive action, the total spatial physical reasonable cost value is determined; the total spatial physical reasonable cost value is used to characterize the unreasonableness of the posture and physical environment of the execution carrier when performing the action.

2. The method according to claim 1, characterized in that, The movement cost condition includes the condition of minimizing the total movement cost indicated by the pending target location information; The step of determining the interaction position corresponding to the at least one interaction action from the at least one candidate position set based on the movement cost condition and the initial interaction position, to obtain the target position information of the execution carrier, includes: Based on the initial interaction position and the interaction position corresponding to the at least one interaction action, a cost objective function is constructed; the cost objective function is used to calculate the total movement cost. With the goal of minimizing the cost objective function, the interaction position corresponding to the at least one interaction action is determined from the at least one candidate position set, thereby obtaining the target position information of the execution carrier.

3. The method according to claim 2, characterized in that, The step of determining the interaction position corresponding to the at least one interaction action from the at least one candidate position set, with the goal of minimizing the cost objective function, to obtain the target position information of the execution carrier, includes: According to the order of the interaction actions of the current task, traverse all candidate position combinations corresponding to each interaction action in turn; During the process of traversing all candidate position combinations corresponding to each interaction action: Calculate the current movement cost corresponding to the candidate interaction positions selected before the current interaction action; Calculate the minimum move cost corresponding to the unselected candidate interaction position after the current interaction action; If the sum of the current move cost and the minimum move cost is greater than the minimum of the current total move cost, the traversal continues from the next candidate interaction position after the previous interaction action.

4. The method according to claim 1, characterized in that, The step of determining the candidate location information corresponding to the execution carrier based on the interaction object type corresponding to the at least one interaction action includes: From the current object instances in the current simulated scene, determine at least one set of candidate instances corresponding to the at least one interactive action; the set of candidate instances includes at least one candidate interactive object instance that matches the type of the interactive object. By utilizing the positional matching relationship between candidate interaction object instances and execution carriers, the interaction position of the execution carrier that matches the position of the candidate interaction object instance is determined, thereby obtaining the candidate position information corresponding to the execution carrier.

5. A simulation implementation device for cloud-native intelligent agent action planning, characterized in that, The apparatus is used to implement the steps of the method according to any one of claims 1 to 4.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images