Agent-based task planning method

Abstract

The application provides an agent-based task planning method, which improves the execution continuity and response efficiency of the agent in a dynamic task target changing scene. First, a natural language description of a task to be executed is obtained and parsed into a hierarchical action primitive sequence, each action primitive being composed of an action type identifier and a semantic attribute slot. Then, a dynamic binding relationship with a target object instance is established when the action primitive is executed, a semantic attribute change event is detected in real time through a target change listening channel, and the binding relationship is switched to a new target object instance without sensing. Finally, the current geometric pose of the next target object is extracted within a transition readiness window, a smooth motion trajectory parameter is generated, and the next action primitive is filled. Through semantic-level action primitive decoupling, dynamic binding switching without sensing, and transition window smooth trajectory generation, the application significantly improves the ability of the agent to complete online switching of the target in a dynamic task environment without shutdown and re-planning.

Description

Technical Field

[0001] This invention relates to the field of task planning, specifically to a task planning method based on intelligent agents. Background Technology

[0002] With the rapid development of artificial intelligence and robotics, intelligent agents are increasingly being used in fields such as intelligent manufacturing, warehousing and logistics, and home services. Task planning, as the core link of autonomous decision-making and execution of intelligent agents, directly determines the operational efficiency and adaptability of intelligent agents in complex and dynamic environments. Traditional task planning methods mostly follow a sequential paradigm of "plan first, then execute." These methods perform well in structured and static environments. However, in practical applications, task objectives often need to be dynamically adjusted according to production scheduling or human-machine interaction instructions. Once traditional methods detect a change in the task objective, they must stop the currently executing trajectory, clear the planned path cache, and rerun the global path planning algorithm based on the position of the new objective, resulting in an interruption of the overall task cycle and a significant decrease in production efficiency.

[0003] Existing technologies have introduced hierarchical task planning frameworks, which decompose complex tasks into multiple sub-tasks and solve them layer by layer, reducing the complexity of single planning to some extent. They also utilize reinforcement learning or imitation learning methods to train agents to learn end-to-end motion strategies in specific scenarios, enabling agents to have a certain degree of environmental generalization ability. However, existing technologies have not fundamentally solved the technical pain point that agents must stop and wait when the task objective changes. Whether it is hierarchical replanning, local path correction, or policy re-inference, all inevitably introduce response delays of hundreds of milliseconds to several seconds. In dynamic operation scenarios that require high-frequency objective switching, the accumulated time loss seriously restricts the overall system's operating efficiency. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes a task planning method based on intelligent agents.

[0005] Therefore, the technical solution adopted by the present invention is as follows:

[0006] An agent-based task planning method, which includes:

[0007] S1: Obtain the natural language description of the task to be executed and parse it into a hierarchical sequence of action primitives for the agent. Each action primitive consists of an action type identifier and a target attribute slot.

[0008] S2: During the execution of the action primitive, a dynamic binding relationship is established with the target object instance, and semantic attribute change events are detected in real time through a preset target change listening channel. Based on the semantic attribute change events, the dynamic binding relationship is seamlessly switched from the target object instance to the new target object instance.

[0009] S3: When the current action primitive is executed to the preset transition ready window, extract the current geometric pose of the target object instance bound to the next action primitive, generate motion trajectory parameters, and fill the motion trajectory parameters into the execution instruction of the next action primitive.

[0010] Furthermore, the hierarchical action primitive sequence is a set of action primitives executed according to the task;

[0011] The action type identifier defines the specific action category that the agent needs to perform;

[0012] The target attribute slot is used in the action primitive to store the semantic attributes of the target object.

[0013] Furthermore, the establishment of the dynamic binding relationship involves binding the semantic attributes in the action primitive to the target object instance in the environment.

[0014] The dynamic binding relationship is a pointer variable pointing to the target object instance;

[0015] Once the dynamic binding relationship is established, the agent will use the current geometric pose of the target object instance as the reference input for the agent's motion control unit during the execution of the action primitive.

[0016] Furthermore, the target change monitoring channel detects task target change instructions and immediately triggers a seamless switching process for dynamic binding relationships upon detecting the change instruction;

[0017] The semantic attribute change event is an instruction indicating that the semantic attribute of the current task target has changed;

[0018] The switching of the dynamic binding relationship is as follows: when a semantic attribute change event occurs, a matching new target object instance is found in the target object instance list according to the changed semantic attribute, and the value of the pointer variable is updated from the memory address of the target object instance to the memory address of the new target object instance.

[0019] Furthermore, during seamless switching, two operations are performed, specifically:

[0020] 1) During seamless handover, a position reference transition period is introduced. The reference pose during the transition period is generated using a weighted interpolation method, represented as:

[0021]

[0022] in, Indicates the time during the handover transition period Motion control reference pose; This represents the geometric pose of the original target object instance at the current moment; This represents the geometric pose of the new target object instance at the current moment; Indicates the time-varying weighting coefficient;

[0023] The decay function of the time-varying weighting coefficient is expressed as a cosine curve:

[0024]

[0025] in, Indicates the moment when the dynamic binding relationship switch occurs; Indicates the preset transition period duration; Indicates the current moment;

[0026] 2) Reset the status of the target change monitoring channel.

[0027] Furthermore, the transition ready window is a time interval defined when the execution progress of the current action primitive reaches a preset proportion range;

[0028] The current geometric pose of the target object instance to which the next action primitive is bound is represented as:

[0029]

[0030] in, , and These represent the three-dimensional position coordinates of the target object instance in the environment coordinate system; , and These respectively represent the target object around , and The rotation angle of the shaft;

[0031] Based on the current geometric pose, motion trajectory parameters are generated using a parameterized motion primitive generator. This parameterized motion primitive generator employs a trajectory planning method based on a fifth-order polynomial, expressed as follows:

[0032]

[0033] in, Represents normalized time The trajectory position at that location, Corresponding to the start time of the next action primitive, The termination time of the next action primitive; The coefficients of the constant term in a fifth-degree polynomial; , , , and These represent the coefficients of the first, second, third, fourth, and fifth degree terms of a fifth-degree polynomial, respectively.

[0034] The six coefficients of the fifth-degree polynomial are determined by six boundary conditions.

[0035] Compared with the prior art, the advantages of the present invention are as follows:

[0036] 1. This invention introduces semantic-level action primitive sequences, transforming task planning from traditional geometric path search into the combination and binding of semantic actions. The action primitives store the semantic attributes of the target object rather than its geometric coordinates, which completely decouples the planning results from the geometric layout of the environment. When the position of the target object changes, the planning level does not need to recalculate, fundamentally avoiding global replanning triggered by the movement of the target.

[0037] 2. This invention achieves the continuity of agent movement when the task target changes through dynamic binding relationship and seamless switching mechanism. After detecting the semantic attribute change event, the system only updates the pointer of the binding relationship at the memory level without interrupting the motion control loop of the current action primitive. The target switching process is completed in microseconds, and the agent does not need to decelerate or stop, thus ensuring the continuity of the task rhythm.

[0038] 3. This invention achieves smooth transitions between action primitives and delayed physical responses to target changes by combining a transition-ready window with a parameterized motion primitive generator. Within the transition window, a smooth trajectory that satisfies the continuity of position, velocity, and acceleration is generated in real time using the geometric pose of the new target. This ensures the correctness of the target change response and avoids abrupt jumps during the motion process, significantly improving the execution efficiency and robustness of the agent in dynamic task environments. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. 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 the task planning method of the present invention;

[0041] Figure 2 This is a schematic diagram of the action primitives of the present invention;

[0042] Figure 3 The flowchart for generating motion trajectory parameters according to the present invention is shown. Detailed Implementation

[0043] To achieve the above objectives, the present invention provides a task planning method based on an intelligent agent, comprising:

[0044] S1: Obtain the natural language description of the task to be executed and parse it into a hierarchical sequence of action primitives for the agent. Each action primitive consists of an action type identifier and a target attribute slot.

[0045] This step transforms easily understandable natural language descriptions into a sequence of structured action instructions that the agent can directly execute. At the same time, it achieves a complete separation between the task description and the specific geometric information of the agent's environment, laying the foundation for subsequent steps where global replanning is not triggered when the task objective changes dynamically.

[0046] Natural language description refers to task instructions expressed by operators in everyday natural language through voice input, text input, or commands from a host computer. Examples include tasks like "grab the red circular workpiece and place it on the workbench in area A" or "screw the screws on the tray into the three holes on the left in sequence." Natural language description is highly abstract and semantically generalizable; it only includes the target semantics and logical relationships of the task, excluding the geometric parameters of the agent's trajectory and the precise coordinates of the target object in the environment.

[0047] After obtaining the natural language description of the task to be executed, a pre-built task-primitive mapping model is invoked to parse the natural language description into a hierarchical sequence of action primitives. The task-primitive mapping model is a pre-built and trained neural network model whose function is to establish a mapping relationship between the semantic space of the natural language description and the set space of action primitives that the agent can execute. The task-primitive mapping model receives the natural language description as input data, transforms the input text into a fixed-dimensional high-dimensional semantic feature vector through an internal semantic encoder, and then decodes the high-dimensional semantic feature vector into a structured hierarchical sequence of action primitives through an action decoder. The training process of the task-primitive mapping model adopts a supervised learning method. The training data consists of a large number of pairs of natural language descriptions and manually annotated standard action primitive sequences. The training objective is to minimize the difference between the model output sequence and the standard annotated sequence.

[0048] A hierarchical action primitive sequence refers to a set of action primitives arranged according to the logical order or temporal constraints of task execution. This sequence has a hierarchical structure, meaning a high-level task may be decomposed into multiple subtasks, each subtask corresponding to an action primitive subsequence. For example, the hierarchical action primitive sequence for the natural language description "grab the red workpiece and place it into box A" is represented as follows:

[0049]

[0050] in, This represents a hierarchical sequence of action primitives; This represents the first action primitive in the sequence; This represents the second action primitive in the sequence; This represents the third action primitive in the sequence; for the specific task mentioned above, the action type identifier of the first action primitive is "move to the vicinity of the target", and the semantic attribute stored in its target attribute slot is "red workpiece"; the action type identifier of the second action primitive is "grab", and the semantic attribute stored in its target attribute slot is "red workpiece"; the action type identifier of the third action primitive is "place", and the semantic attribute stored in its target attribute slot is "box A".

[0051] Each action primitive consists of two core components: an action type identifier and a target attribute slot. The action type identifier defines the specific action category that the agent needs to perform. This action category comes from a predefined action library, which contains all atomic actions that the agent can perform at the physical level, such as "move to the vicinity of the target," "grab," "place," "push," "rotate," "press," and "release." Each action type identifier corresponds to a pre-defined motion control strategy, which encapsulates the underlying motion planning algorithm and control parameter templates required to complete the action.

[0052] The target attribute slot is a data field in action primitives used to store descriptive information about the target object. It stores the semantic attributes of the target object, not its geometric coordinates. Semantic attributes refer to the descriptive features of the target object that distinguish it within the task context, such as "red workpiece," "blue cup," "workbench in area A," "largest circular object," and "button on the left." These semantic attributes describe abstract features such as the target object's category, color, shape, and relative position, rather than its specific numerical coordinates in the environment coordinate system. In contrast, geometric coordinates refer to the precise spatial location description of the target object in the environment coordinate system.

[0053] The fundamental reason for designing target attribute slots to store semantic attributes rather than geometric coordinates is that semantic attributes have a natural decoupling characteristic from the geometric layout of the environment. No matter how the target object moves or changes in its specific position in the environment, its inherent semantic attributes remain unchanged. For example, a "red circular workpiece" will not change its semantic attribute "red circular workpiece" regardless of whether it is placed at the beginning, middle, or end of a conveyor belt. Therefore, the action primitive sequence composed of semantic attributes does not depend on specific target coordinate values. When the position of the target object moves due to dynamic changes in the environment, the action primitive sequence itself does not need to be modified or replanned. It is only necessary to establish a dynamic binding relationship between the semantic attributes in the slot and the actual target object instance in the environment through perception in the subsequent execution stage. This design transforms task planning from the "static geometric path preset" paradigm in traditional methods to the "dynamic semantic intent preservation" paradigm. This is the technical foundation for achieving seamless switching of task objectives and avoiding global replanning.

[0054] S2: During the execution of the action primitive, a dynamic binding relationship is established between the target object instance and the target object instance. Semantic attribute change events are detected in real time through a preset target change listening channel. Based on the semantic attribute change events, the dynamic binding relationship is seamlessly switched from the target object instance to the new target object instance.

[0055] During the process of the agent executing a task according to the hierarchical action primitive sequence generated in step one, when the upper-level scheduler or operator issues a task target change instruction, the binding relationship between the current action primitive and the target object is dynamically redirected. Moreover, this redirection process does not cause the agent's motion state to be interrupted, thus completely avoiding the problem of downtime and replanning that must be triggered due to target changes in traditional task planning methods.

[0056] Dynamic binding refers to the process of establishing a one-to-one correspondence between the semantic attribute descriptions stored in the target attribute slots of action primitives and the target object instances actually observed by the agent in the current environment through perception. Since the action primitive sequence generated in step one only stores the semantic attributes of the target objects rather than their geometric coordinates, the abstract semantic attributes need to be translated into concrete target objects in the environment before the action primitives are actually executed. This translation process is the establishment of dynamic binding.

[0057] In this embodiment, when the agent begins to execute the first action primitive in the hierarchical action primitive sequence... At this time, the action type identifier of the action primitive is "move to the vicinity of the target", and the semantic attribute stored in the target attribute slot is "red workpiece". The agent's perception unit scans the current environment and uses a pre-set target detection and semantic recognition model to identify all target object instances and their corresponding semantic attributes in the environment. Assuming there are three workpiece instances in the environment, with semantic attributes of "red workpiece", "blue workpiece" and "green workpiece" respectively, the agent will use the action primitive... The semantic attribute "red workpiece" of the target attribute slot is matched with red workpiece instances in the environment to establish a dynamic binding relationship. The formal representation of this dynamic binding relationship is as follows:

[0058]

[0059] in, This represents the first hierarchical action primitive sequence. Action primitives; This represents the first [object] in the environment identified by the intelligent agent's perception unit. There are 1 target object instance, and each target object instance has a unique instance identifier; This represents a dynamically bound function, which takes an action primitive as input and outputs the target object instance to which it is bound.

[0060] After the dynamic binding relationship is established, the agent executes the action primitive. During the process, the target object instance will be used. The current geometric pose serves as the reference input for the motion control unit. Since the target object in the environment may change position due to conveyor belt movement, human movement, or operations by other intelligent agents, the target object instance... The geometric pose changes dynamically over time. The agent's perception unit continuously tracks the current pose of the bound target object instance at a fixed frequency and transmits the latest pose information to the motion control unit in real time to ensure that the motion trajectory can adapt to the movement of the target.

[0061] The target change monitoring channel is an asynchronous monitoring mechanism that runs independently of the motion control main thread. Its function is to detect target change commands from upper-layer task scheduling or human-computer interaction interfaces in real time, and immediately trigger a dynamic binding relationship switching process upon detection. The target change monitoring channel periodically checks the command queue in the agent's control unit at a preset polling frequency, typically set to 100 Hz (every 10 milliseconds), to ensure that the response latency to target change commands is controlled within milliseconds.

[0062] A semantic attribute change event refers to an instruction signal issued by the upper-level system to indicate that the semantic attribute of the current task target has changed. For example, in a manufacturing scenario, when the type of workpiece upstream on the assembly line changes from "red workpiece" to "blue workpiece", a semantic attribute change instruction is sent to the intelligent agent. This instruction contains the changed semantic attribute value "blue workpiece".

[0063] When the target change monitoring channel detects a semantic attribute change event in the instruction queue, the monitoring channel immediately parses the event content, extracts the changed semantic attribute value, and triggers the dynamic binding relationship switching process.

[0064] Seamless switching refers to the process of dynamically switching the binding relationship from a target object instance to a new target object instance without causing any interruption, deceleration, or pause in the agent's current motion state. In traditional task planning methods, when the task objective changes, it is usually necessary to first suspend the currently executing motion trajectory, clear the planned path cache, rerun the global path planning algorithm based on the new target's position to generate a new motion trajectory, and then restart motion control. This complete process typically takes between hundreds of milliseconds and several seconds. During this period, the agent must be in a stopped and waiting state, leading to interruptions in the overall task cycle and a decrease in productivity.

[0065] The seamless switching mechanism in this step achieves zero interference with motion execution during the switching process. Specifically, the execution of the current action primitive depends only on the real-time geometric pose of the target object instance pointed to by the dynamic binding relationship. The update operation of the dynamic binding relationship itself is only a reassignment of pointers or object references in memory at the software implementation level. The dynamic binding relationship is implemented in the program as a pointer variable pointing to the target object instance. When a semantic attribute change occurs, the system first searches for a new target object instance that matches the changed semantic attribute value in the target object instance list maintained by the environment perception unit. Then, the value of the pointer variable is updated from the memory address of the original target object instance to the memory address of the new target object instance. This pointer assignment operation can be completed in just a few CPU clock cycles in modern computer architectures, with a time overhead in the nanosecond to microsecond range, which is much smaller than the time constant of the intelligent agent motion control system (usually in the millisecond range). Therefore, from the perspective of the motion control thread, the pose information flow of the target object is not interrupted; only the source object of the pose data has changed. This seamless switching process is represented as follows:

[0066]

[0067] in, This indicates the dynamic binding relationship corresponding to the currently executing action primitive; This represents the assignment operator, used to update the binding relationship; Represents a new target object instance that matches the changed semantic attributes;

[0068] To ensure the continuity and smoothness of motion control after a seamless transition, the following two auxiliary operations are executed simultaneously at the moment the dynamic binding relationship is switched.

[0069] First, a smooth transition is performed on the motion control reference input under the new binding relationship. Since the geometric poses of the original and new target object instances in the environment may differ significantly, if the reference pose of the motion control unit undergoes a step change during the switching instant, it will cause a sudden change in the velocity or torque commands output by the motion control algorithm, leading to jitter or overshoot in the agent. To address this issue, a brief position reference transition period is introduced during dynamic binding relationship switching. The reference pose during the transition period is generated using weighted interpolation, represented as follows:

[0070]

[0071] in, Indicates the time during the handover transition period Motion control reference pose; This represents the geometric pose of the original target object instance at the current moment; This represents the geometric pose of the new target object instance at the current moment; This represents the time-varying weighting coefficient, whose value smoothly decays from 1 to 0 during the transition period. The decay function uses a cosine curve and is expressed as follows:

[0072]

[0073] in, Indicates the moment when the dynamic binding relationship switch occurs; This indicates the preset transition period duration, which is set to 200 milliseconds in this embodiment; Indicates the current moment; satisfies ;when hour, The reference pose is exactly equal to the original target pose; when hour, The reference pose is completely equal to the new target pose; during the transition period, the reference pose smoothly transitions between the two target poses, effectively avoiding abrupt changes in the reference input.

[0074] Second, the state of the target change monitoring channel is reset. After a seamless switch of dynamic binding relationship is completed, the target change monitoring channel immediately clears the currently processed semantic attribute change events and re-enters the monitoring state, ready to receive the next possible target change instruction. This continuous monitoring and instant response mechanism enables the agent to adapt to complex operation scenarios where task objectives change frequently and dynamically. For example, on a multi-variety mixed production line, the agent may need to process multiple workpieces with different semantic attributes continuously in a short period of time, and the seamless switching mechanism ensures that each target switch will not cause a pause in task execution.

[0075] Step two enables dynamic binding of target objects and seamless switching when targets change during task execution. This step, together with the semantic-level action primitive sequence formed in step one, constitutes a new task planning paradigm that separates planning and execution and decouples semantics and geometry. This provides the correct target pose input and continuous motion state context for the smooth generation of motion trajectories based on transition ready windows in step three.

[0076] S3: When the current action primitive reaches the preset transition ready window, extract the current geometric pose of the target object instance bound to the next action primitive, generate motion trajectory parameters, and fill the motion trajectory parameters into the execution instructions of the next action primitive.

[0077] As the agent nears completion of the current action primitive, the current geometric pose of the new target object instance pointed to by the dynamically bound relationship updated in step two is used to generate the motion trajectory parameters required for the next action primitive in real time, achieving a smooth transition between two adjacent action primitives. This step works in conjunction with the seamless switching mechanism in step two to transform the target semantic switching completed in step two into an actual physical motion response, enabling the agent to smoothly approach and operate on the new task target without downtime or global replanning.

[0078] The transition ready window refers to a specific time interval defined when the execution progress of the current action primitive reaches a preset proportion. Within this time interval, the main motion phase of the current action primitive is basically completed, but the motion has not completely terminated. The agent's motion state is in a stage suitable for planning and preparing for the trajectory switching of the next action primitive. The purpose of setting the transition ready window is to reserve sufficient planning time for the trajectory generation of the next action primitive in the time dimension, while ensuring that the trajectory switching occurs at the time of natural transition of the motion state, avoiding motion discontinuity or control command conflicts caused by switching too early or too late.

[0079] The trigger condition for the transition ready window is based on the current execution progress percentage of the action primitive, and is expressed as follows:

[0080]

[0081] in, Indicates the execution time of the current action primitive; Indicates the estimated total execution time of the current action primitive; This represents the initial scaling factor for the transition-ready window, with a value range of [value missing]. The specific value is configured according to the action type and task requirements. For example, for an action primitive of the type "move to the vicinity of the target", A value of 0.85 indicates that the agent enters the transition ready window when it is only 15% away from the target.

[0082] When the current action primitive is detected to enter the transition ready window, the transition preparation process is immediately executed. First, the current geometric pose of the target object instance to which the next action primitive is bound is extracted from the currently valid dynamic binding relationship, including the 3D position coordinates and 3D pose angle of the target object in the environment coordinate system, represented as:

[0083]

[0084] in, , and These represent the three-dimensional position coordinates of the target object instance in the environment coordinate system; , and These respectively represent the target object around , and The rotation angle of the shaft;

[0085] Subsequently, the parameterized motion primitive generator is invoked to generate the motion trajectory parameters of the next action primitive in real time based on the extracted geometric pose of the target object. The parameterized motion primitive generator is a motion planning unit whose core function is to generate a smooth trajectory that satisfies kinematic and dynamic constraints based on the target pose and the current agent state.

[0086] The parameterized motion primitive generator employs a trajectory planning method based on quintic polynomials. Quintic polynomials ensure the continuity of the trajectory at the position, velocity, and acceleration levels, thereby guaranteeing smooth transitions between motion primitives. The trajectory is parameterized as follows:

[0087]

[0088] in, Represents normalized time The trajectory position at that location, , representing the normalized time variable, Corresponding to the start time of the next action primitive, The termination time of the next action primitive; The coefficients of the constant term in a fifth-degree polynomial; , , , and These represent the coefficients of the first, second, third, fourth, and fifth degree terms of a fifth-degree polynomial, respectively.

[0089] The six coefficients of the fifth-degree polynomial are uniquely determined by six boundary conditions. These six boundary conditions specify the position, velocity, and acceleration values ​​of the trajectory at the initial and final moments, respectively. Specifically, the boundary conditions are defined as follows:

[0090] Starting position conditions: , indicating the normalized time At that time, the trajectory position is equal to the starting position. ;

[0091] Initial speed conditions: , indicating the normalized time At that time, the trajectory velocity is equal to the initial velocity. ;

[0092] Initial acceleration conditions: , indicating the normalized time At that time, the trajectory acceleration is equal to the initial acceleration. ;

[0093] Termination position condition: , indicating the normalized time At that time, the trajectory position is equal to the termination position. ;

[0094] Termination speed condition: , indicating the normalized time At that time, the trajectory velocity equals the termination velocity. ;

[0095] Termination of acceleration condition: , indicating the normalized time At that time, the trajectory acceleration equals the termination acceleration. ;

[0096] The initial position, initial velocity, and initial acceleration are determined by the motion state at the end of the current action primitive to ensure state continuity; the final position is determined by the geometric pose of the target object; and the final velocity and final acceleration are set according to the requirements of the next action primitive, such as a grabbing action requiring zero velocity upon arrival.

[0097] Based on the above six boundary conditions, by solving the system of linear equations, the analytical expression for the coefficients of the fifth-degree polynomial is obtained, which is expressed as:

[0098]

[0099]

[0100]

[0101]

[0102]

[0103]

[0104] The derivation of the above coefficient calculation formula is based on the mathematical principle that the value of a polynomial function and its derivative at the boundary points is equal to the given boundary conditions. For example, Because when hour, Therefore Directly equal to the starting position ; Because the trajectory velocity function is right The derivative, i.e. ,when hour, Therefore equal to the initial velocity ; Because the trajectory acceleration function is right The derivative; , and The expression is then passed through Substituting the position, velocity, and acceleration conditions at time into the polynomial and its derivative equations, and solving them simultaneously, we obtain the solution.

[0105] After calculating the fifth-order polynomial coefficients using the above formula, the parameterized motion primitive generator can generate a complete trajectory from the current state to the target pose and fill the trajectory parameters into the execution instruction of the next action primitive. The agent seamlessly switches to the execution state of the next action primitive the instant the current action primitive is executed and continues to move according to the generated trajectory parameters. Since the state continuity constraint has been considered during trajectory planning, the transition between action primitives is smooth and there are no sudden changes in velocity or acceleration.

[0106] The agent-based task planning method proposed in this invention decouples the task description from the environmental geometry by parsing the natural language task description into a semantically hierarchical sequence of action primitives. During execution, a dynamic binding relationship is established between the action primitives and the target object instance, and a target change monitoring channel is used to achieve seamless switching of the binding relationship. At the same time, when the current action primitive is executed to the preset transition ready window, motion trajectory parameters that meet the continuity constraints are generated in real time based on the geometric pose of the new target object. This allows the agent to smoothly transition to the new task target without stopping and replanning when the task target changes dynamically.

[0107] In summary, this invention fundamentally breaks through the technical bottleneck of the traditional task planning method, where target changes must trigger global replanning under the sequential paradigm of planning before execution. It transforms task planning from static geometric path pre-setting to dynamic semantic intent maintenance, realizing the unity of online switching of task targets and continuous maintenance of motion execution. This significantly improves the operational efficiency and robustness of intelligent agents in scenarios with frequent target changes, such as mixed-flow production and dynamic picking.

[0108] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A task planning method based on intelligent agents, characterized in that, The method includes: S1: Obtain the natural language description of the task to be executed and parse it into a hierarchical sequence of action primitives for the agent. Each action primitive consists of an action type identifier and a target attribute slot. S2: During the execution of the action primitive, a dynamic binding relationship is established with the target object instance, and semantic attribute change events are detected in real time through a preset target change listening channel. Based on the semantic attribute change events, the dynamic binding relationship is seamlessly switched from the target object instance to the new target object instance. S3: When the current action primitive is executed to the preset transition ready window, extract the current geometric pose of the target object instance bound to the next action primitive, generate motion trajectory parameters, and fill the motion trajectory parameters into the execution instruction of the next action primitive.

2. The agent-based task planning method according to claim 1, characterized in that, The hierarchical sequence of action primitives is a set of action primitives executed according to the task. The action type identifier defines the specific action category that the agent needs to perform; The target attribute slot is used in the action primitive to store the semantic attributes of the target object.

3. The agent-based task planning method according to claim 2, characterized in that, The establishment of the dynamic binding relationship involves binding the semantic attributes in the action primitive to the target object instance in the environment. The dynamic binding relationship is a pointer variable pointing to the target object instance; Once the dynamic binding relationship is established, the agent will use the current geometric pose of the target object instance as the reference input for the agent's motion control unit during the execution of the action primitive.

4. The agent-based task planning method according to claim 3, characterized in that, The target change monitoring channel detects task target change instructions and immediately triggers a seamless switching process for dynamic binding relationships upon detection of the change instruction. The semantic attribute change event is an instruction indicating that the semantic attribute of the current task target has changed; The switching of the dynamic binding relationship is as follows: when a semantic attribute change event occurs, a matching new target object instance is found in the target object instance list according to the changed semantic attribute, and the value of the pointer variable is updated from the memory address of the target object instance to the memory address of the new target object instance.

5. The agent-based task planning method according to claim 4, characterized in that, During seamless switching, two operations are performed, specifically: 1) During seamless handover, a position reference transition period is introduced. The reference pose during the transition period is generated using a weighted interpolation method, represented as: in, Indicates the time during the handover transition period Motion control reference pose; This represents the geometric pose of the original target object instance at the current moment; This represents the geometric pose of the new target object instance at the current moment; Indicates the time-varying weighting coefficient; The decay function of the time-varying weighting coefficient is expressed as a cosine curve: in, Indicates the moment when the dynamic binding relationship switch occurs; Indicates the preset transition period duration; Indicates the current moment; 2) Reset the status of the target change monitoring channel.

6. The agent-based task planning method according to claim 5, characterized in that, The transition ready window is a time interval defined when the execution progress of the current action primitive reaches a preset proportion range; The current geometric pose of the target object instance to which the next action primitive is bound is represented as: in, , and These represent the three-dimensional position coordinates of the target object instance in the environment coordinate system; , and These respectively represent the target object around , and The rotation angle of the shaft; Based on the current geometric pose, motion trajectory parameters are generated using a parameterized motion primitive generator. This parameterized motion primitive generator employs a trajectory planning method based on a fifth-order polynomial, expressed as follows: in, Represents normalized time The trajectory position at that location, Corresponding to the start time of the next action primitive, The termination time of the next action primitive; The coefficients of the constant term in a fifth-degree polynomial; , , , and These represent the coefficients of the first, second, third, fourth, and fifth degree terms of a fifth-degree polynomial, respectively. The six coefficients of the fifth-degree polynomial are determined by six boundary conditions.

Images