An agent task dynamic allocation and cooperation system

Abstract

This invention discloses a dynamic task allocation and collaboration system based on intelligent agents, belonging to the field of intelligent agent collaborative control technology. It includes a task parsing module, a task decomposition module, a dynamic allocation module, an interaction sensing module, a conflict prediction module, and a collaborative adjustment module. The task parsing module acquires the target transport task parameters and parses spatial and temporal constraints. The dynamic allocation module dynamically selects the main and auxiliary intelligent agents from a heterogeneous intelligent agent set based on their current pose, load capacity, and communication latency threshold to execute each sub-task, forming a collaborative intelligent agent group. The interaction sensing module collects physical interaction force data and relative pose data between intelligent agents in real time. The conflict prediction module constructs a conflict probability field within a shared workspace based on the above data. When the field strength characteristic value exceeds a first preset threshold, the collaborative adjustment module inserts a collaborative safety adjustment phase into subsequent unexecuted sub-tasks to adjust the execution sequence or motion trajectory of the main and auxiliary intelligent agents.

Description

Technical Field

[0001] This invention relates to the field of intelligent agent cooperative control technology, specifically to an intelligent agent-based dynamic task allocation and cooperation system. Background Technology

[0002] In multi-agent collaborative material handling operations, complex handling tasks typically need to be broken down into multiple sub-tasks based on spatial and temporal constraints, and agents with different capabilities are assigned to complete them collaboratively. Existing task allocation technologies mainly rely on static or quasi-static information such as the agent's current position and load capacity for one-time assignment, lacking real-time fusion analysis of the physical interaction state and relative pose changes between agents during task execution. When multiple agents share a narrow working space and simultaneously bear the handling load, simple predictions based solely on position coordinates and velocity vectors are insufficient to accurately assess the collision risks caused by force imbalances and clamping force fluctuations. When potential interference is detected, existing technologies often resort to emergency stopping or single obstacle avoidance, failing to proactively coordinate the behavior of multiple agents while maintaining task execution sequence and overall progress. This makes it easy for agents to slip due to mutual squeezing forces or trajectory conflicts in multi-agent collaborative material handling with strict spatial constraints and tight time windows, leading to cargo slippage, agent damage, or task timeouts, thus restricting both system safety and task completion efficiency.

[0003] For multiple subtasks with execution sequence dependencies, existing solutions cannot utilize physical interaction force data and relative pose data to generate a quantified conflict risk distribution in real time within a shared workspace. Nor can they proactively insert a collaborative safety adjustment phase into the subsequent unexecuted subtask processes when the conflict risk exceeds the safety boundary, dynamically shifting the execution sequence or replanning the motion trajectory online. Therefore, it is urgent to address how to construct a potential field reflecting the probability of force-potential coupling conflicts through continuous perception of interaction forces and poses during close multi-agent collaboration, and how to automatically intervene in the execution progress and path of subtasks when the conflict probability is too high, in order to eliminate spatiotemporal overlapping conflicts without excessively sacrificing overall transport efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a dynamic task allocation and collaboration system based on intelligent agents. This system can construct a conflict probability field based on real-time collected physical interaction force and relative pose data during multi-agent collaborative handling, forming a quantitative description of the force-position coupling conflict risk. When the field strength characteristics of the conflict probability field exceed a preset threshold, a collaborative safety adjustment phase is proactively inserted into the unexecuted sub-tasks. By adjusting the time window and motion trajectory of the sub-tasks, conflict resolution is achieved, thereby improving the safety level and task completion efficiency of heterogeneous multi-agent collaborative handling under complex spatial and temporal constraints.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] This invention provides a dynamic task allocation and collaboration system based on intelligent agents, including a task parsing module, a task decomposition module, a dynamic allocation module, an interaction perception module, a conflict prediction module, and a collaborative adjustment module. The task parsing module acquires the task parameters of the target transportation task and parses the spatial and temporal constraints within the target transportation task, enabling the system to accurately understand the rigid requirements of the task in the spatial and temporal dimensions. The task decomposition module decomposes the target transportation task into multiple sub-tasks with execution order dependencies based on the spatial and temporal constraints. Preferably, the task decomposition module discretizes the transportation path into multiple path segments based on the spatial constraints, marks the starting and ending spatial coordinates of each path segment, and assigns a maximum allowable passage time to each path segment based on the temporal constraints. A path segment and its corresponding maximum allowable passage time are determined as an independent sub-task, and the execution order dependencies between sub-tasks are determined based on the adjacency relationship between the starting and ending spatial coordinates. This decomposition method also calculates the required load-bearing capacity of each sub-task based on the cargo weight, path slope, and friction coefficient of each path segment, and allocates the maximum response delay to the sub-task based on the maximum allowable passage time. It can even further calculate the minimum travel speed requirement of each sub-task as a dynamic constraint, so that the generated sub-tasks have clear operational boundaries and quantifiable performance indicators, laying the foundation for accurate matching of heterogeneous intelligent agents.

[0007] The dynamic allocation module dynamically selects the primary and secondary agents from the heterogeneous agent set to execute each sub-task based on the current pose, load capacity, and communication latency threshold of each agent, forming a collaborative agent group for the target transport task. As a preferred embodiment of the invention, the dynamic allocation module obtains the navigation distance between the current pose of each agent and the starting point of the sub-task. Based on the ratio of navigation distance, load capacity to the required carrying capacity of the sub-task, and the difference between the communication latency threshold and the maximum allowable response latency of the sub-task, the module calculates the adaptation weight coefficient for each agent to the sub-task. The agent with the highest adaptation weight coefficient is determined as the primary agent, and the agent with the second highest adaptation weight coefficient and whose communication protocol is compatible with the primary agent is determined as the secondary agent. This ensures that the selected agent group is highly adapted to the task requirements in terms of spatial accessibility, carrying capacity, and real-time communication.

[0008] The interactive perception module collects physical interaction force data and relative pose data between the collaborative intelligent agents in real time during the execution of the sub-tasks. Specifically, a six-dimensional force sensor installed at the end of each intelligent agent collects physical interaction force data between the main intelligent agent and the auxiliary intelligent agent, as well as between the intelligent agent and the cargo. The three-dimensional spatial coordinates and orientation angles of each intelligent agent are obtained in real time through an ultra-wideband positioning module or a visual odometry. After coordinate system transformation, the relative distance and relative attitude angle between each pair of intelligent agents are calculated as relative pose data, thereby providing fine-grained multi-source perception information for conflict prediction.

[0009] The conflict prediction module constructs a conflict probability field for the collaborative intelligent agent group within the shared workspace based on the physical interaction force data and the relative pose data. In a preferred embodiment, the predicted motion trajectory of each agent in the next control cycle is predicted based on its current movement velocity vector and relative pose data. The width expansion coefficient of the predicted motion trajectory is corrected based on the normal and tangential force components in the physical interaction force data. A single-agent potential field is generated for each agent using the centerline of the predicted motion trajectory as the field source and the width expansion coefficient as the field radius. All single-agent potential fields are then superimposed, and the region where the superimposed potential field value is higher than a second preset threshold is marked as a high-probability region of the conflict probability field. The width expansion coefficient can increase linearly with the magnitude of the tangential force component, and the upper limit of the increase does not exceed half of the minimum safe obstacle avoidance distance between the main agent and the auxiliary agents, ensuring that the potential field radius accurately reflects the uncertain motion range caused by fluctuations in interaction forces.

[0010] When the field strength characteristic value of the conflict probability field exceeds a first preset threshold, the collaborative adjustment module inserts a collaborative safety adjustment phase into the subsequent unexecuted subtasks to adjust the execution sequence or movement trajectory of the main agent and the auxiliary agent. As a specific technical solution of this invention, the field strength characteristic value can be the weighted root mean square value of the potential field values ​​of all grid points in the conflict probability field, or it can be the average potential field value within each connected region extracted from the high-probability region. When the field strength characteristic value of any connected region exceeds the first preset threshold, the conflict agent combination corresponding to that connected region is identified. Based on the priority of the subtasks currently being executed by each agent in the conflict agent combination, the maximum allowed passage time corresponding to the lower-priority subtasks is temporarily extended, and the maximum allowed passage time corresponding to the higher-priority subtasks is correspondingly compressed, thereby adjusting the execution sequence. If the field strength characteristic value still exceeds the first preset threshold after synchronously updating the conflict probability field, a further local replanning of the movement trajectory of the conflict agent combination is triggered, thereby reliably resolving the operational conflict.

[0011] Furthermore, the system may also include a collaborative obstacle avoidance control module, used to acquire the contour point cloud and velocity vector of dynamic obstacles in front of the main agent, calculate the first remaining time for the main agent to collide with the dynamic obstacle at the current speed, and when the first remaining time is less than a safe time threshold, generate a lateral avoidance offset vector and broadcast it to the auxiliary agent, controlling the auxiliary agent to synchronously superimpose the lateral avoidance offset vector while maintaining a constant relative attitude angle with the main agent, forming a formation overall offset motion to achieve safe obstacle avoidance in a cooperative state. The system may also include a collaborative load balancing module, used to monitor the first current value of the main agent's drive motor and the second current value of the auxiliary agent's drive motor in real time, calculate the instantaneous deviation, and generate a torque redistribution command when the deviation exceeds the load balancing threshold, adjusting the output torque of the auxiliary agent through a low-latency communication link so that the first current value and the second current value tend to be consistent in the next control cycle, ensuring load balance during the cooperative handling process and avoiding overload of a single agent. The system may also include a local autonomous collaboration module. When a communication interruption is detected between the primary and secondary agents, the primary agent switches to a local predictive control mode. Based on the motion state parameters of the most recently received secondary agent, it calculates the motion trajectory of the secondary agent during the interruption and reduces its own movement speed, increasing the safety protection gap between itself and the motion prediction trajectory. After communication is re-established, the trajectory logs during the interruption are exchanged, and the default value of the safety protection gap is dynamically updated based on the maximum approach distance, thus maintaining basic collaborative safety even in communication anomalies. The system may also include a collaboration efficiency evaluation module. After completing the target transport task, it records the total time, number of conflicts, and average conflict duration, calculates the overall collaboration efficiency coefficient, and stores it in a historical task database along with the task parameters. This allows the system to retrieve the allocation strategy with the highest efficiency coefficient as a reference for the initial allocation scheme when receiving subsequent transport tasks with similarity exceeding a threshold, enabling the system to have self-learning and strategy optimization capabilities.

[0012] Through the coordinated operation of the above modules, this invention adaptively decomposes complex handling tasks into ordered sub-tasks, dynamically assembles optimal cooperative intelligent agent groups based on spatial and temporal constraints, and constructs a conflict probability field in real time based on force perception and pose information within a shared work space, triggering dual adjustments to the execution sequence and trajectory. With the assistance of mechanisms such as formation obstacle avoidance, load balancing, autonomous cooperation during communication interruptions, and performance evaluation, it can achieve safe and efficient cooperation among multiple heterogeneous intelligent agents in dynamic and uncertain environments, significantly improving the flexibility, robustness, and overall operational efficiency of task execution.

[0013] The beneficial effects of this invention are:

[0014] Physical interaction force data between the master and auxiliary agents, and between the agents and the cargo, is collected using distributed six-dimensional force sensors. Simultaneously, ultra-wideband positioning or visual odometry is used to acquire the three-dimensional spatial coordinates and orientation angles of each agent. After coordinate system integration, relative distance and relative attitude angles are calculated as relative pose data. Based on this, for each agent, the predicted motion trajectory for the next control cycle is predicted by combining its current movement velocity vector and relative pose. Normal and tangential force components are extracted from the physical interaction force data. The amplitude of the tangential force component is used to linearly correct the width expansion coefficient of the predicted motion trajectory. The increase in this expansion coefficient does not exceed half of the minimum safe obstacle avoidance distance between the master and auxiliary agents. A single-agent potential field is generated using the centerline of the predicted motion trajectory as the field source and the expanded width as the field radius. The single-agent potential fields of all agents are superimposed within a shared working space. Regions with superimposed potential field values ​​higher than a second preset threshold are marked as high-probability regions of the conflict probability field. Connected components are then extracted, and the average potential field value within the connected components is calculated as the field strength feature value. Since the width expansion coefficient is directly related to the tangential interaction force, when a large tangential force is generated by pushing or friction between agents, the area occupied by the predicted trajectory is dynamically widened. The conflict probability field will generate a higher potential field value in the direction of intense force interaction, transforming the abstract internal force state into an external collision risk spatial distribution. It can identify potential collision areas induced by torque imbalance before the agents actually make contact or go beyond the distance, overcoming the defect that relying solely on position and velocity prediction cannot reflect the risk of force interaction, making conflict prediction more sensitive and accurate to real physical processes.

[0015] When the field strength characteristic value of any connected domain exceeds a first preset threshold, the conflicting agent combination corresponding to that connected domain is identified. Based on the priority of the subtasks currently being executed by each conflicting agent, the maximum allowable passage time corresponding to the lower-priority subtask is temporarily extended, while the maximum allowable passage time of the higher-priority subtask is correspondingly compressed. This widens the temporal misalignment interval between the two subtasks, achieving spatial conflict resolution through flexible adjustment of execution timing. If the field strength characteristic value still exceeds the limit after updating the conflict probability field, a further local replanning of the movement trajectory of the conflicting agent combination is triggered to eliminate the remaining conflict points from a spatial dimension. By embedding a collaborative safety adjustment phase in the unexecuted subtasks associated with the conflict, the spatiotemporal occupancy windows of agents can be proactively staggered without interrupting the entire transport task. This avoids situations where two agents block each other or simultaneously compete for the same spatial area at narrow passages or intersections. High-risk conflicts can be resolved locally without global task reordering, maintaining the rapid progress of high-priority subtasks while preventing low-priority subtasks from being delayed indefinitely. This mechanism, which combines timing adjustment with trajectory replanning, is suitable for complex and dynamic environments. It can effectively reduce the stagnation time and collision probability during the transportation process, and improve the smoothness of multi-agent collaboration and the reliability of tasks under strict time constraints. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a schematic diagram of a system based on dynamic task allocation and collaboration of intelligent agents;

[0018] Figure 2 It is a task breakdown flowchart;

[0019] Figure 3 This is a flowchart of the dynamic allocation module generating collaborative intelligent agent groups; Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See Figure 1This invention provides a dynamic task allocation and collaboration system based on intelligent agents, comprising: a task parsing module, which acquires task parameters of a target transportation task and parses the spatial and temporal constraints in the target transportation task; a task decomposition module, which decomposes the target transportation task into multiple sub-tasks with execution order dependencies based on the spatial and temporal constraints; a dynamic allocation module, which dynamically selects a master agent and an auxiliary agent to execute each sub-task from a heterogeneous set of intelligent agents, based on the current pose, load capacity, and communication latency threshold of each agent, forming a collaborative intelligent agent group for the target transportation task; an interaction perception module, which collects physical interaction force data and relative pose data between the agents in real time during the execution of the sub-tasks by the collaborative intelligent agent group; a conflict prediction module, which constructs a conflict probability field of the collaborative intelligent agent group in a shared workspace based on the physical interaction force data and the relative pose data; and a collaborative adjustment module, which inserts a collaborative safety adjustment phase into the subsequent unexecuted sub-tasks when the field strength characteristic value of the conflict probability field exceeds a first preset threshold, adjusting the execution timing or movement trajectory of the master agent and the auxiliary agent.

[0022] In specific implementation, please refer to Figure 2 The task decomposition module acquires the spatial and temporal constraints output by the task parsing module. Spatial constraints include the geometric boundaries of the transport path, passable width, height limits, and static obstacle distribution information. Based on these spatial constraints, the task decomposition module discretizes the transport path into multiple path segments. The discretization process employs equal-chord length sampling or adaptive sampling based on path curvature, ensuring that the curvature variation within each path segment does not exceed a preset curvature variation threshold. For each path segment, the task decomposition module marks the starting and ending spatial coordinates, both represented in a three-dimensional Cartesian coordinate system, and simultaneously records the ground material properties of the path segment.

[0023] The task decomposition module assigns a maximum allowable transit time to each path segment based on time constraints. These constraints include the total completion time for the target transport task and segmented time requirements. The module allocates the total time to each path segment according to its proportion to the total transport path length, thus obtaining the maximum allowable transit time for each segment. For path segments with special time requirements, the maximum allowable transit time is adjusted based on a weighted average, with the weighting factor set according to the urgency of the special time requirements.

[0024] The task decomposition module defines each path segment and its corresponding maximum allowed passage time as an independent subtask. The subtask's task description includes the path segment identifier, starting point spatial coordinates, ending point spatial coordinates, maximum allowed passage time, and the ground material properties of the path segment. The module determines the execution order dependencies between subtasks based on the adjacency of their starting and ending point spatial coordinates. When the Euclidean distance between the ending point spatial coordinates of one subtask and the starting point spatial coordinates of another subtask is less than a preset adjacency distance threshold, these two subtasks are considered to have a sequential dependency, with the subtask ending after the preceding subtask being marked as the successor subtask. By traversing all subtasks, a directed acyclic graph (DAG) is constructed, with directed edges representing execution order dependencies.

[0025] For each subtask, the task breakdown module calculates the required load-bearing capacity based on the cargo weight, path slope, and surface friction coefficient of the corresponding path segment. The cargo weight is obtained from the task parameters of the target handling task. The path slope is determined by the ratio of the elevation difference between the starting and ending spatial coordinates of the path segment to its horizontal projected length. The surface friction coefficient is retrieved from a pre-built path material friction coefficient database, which stores calibration values ​​for friction coefficients corresponding to different ground material types. The formula for calculating the required load-bearing capacity is:

[0026]

[0027] in, Indicates the first The required carrying capacity of each subtask corresponding to a path segment, in Newtons; Indicates the first The weight of goods to be transported for each subtask along its corresponding path segment, in kilograms; Represents the gravitational acceleration constant, with a value of ; Indicates the first The slope angle of each subtask's path segment relative to the horizontal plane, in radians, is calculated from the arctangent of the elevation difference between the starting points of the path segment and the horizontal projection length. Indicates the first The surface friction coefficient values ​​for each subtask's corresponding path segment are obtained from the path material friction coefficient database based on the ground material attribute index of the path segment. The range of the surface friction coefficient values ​​is as follows: to The required load-bearing capacity mentioned above takes into account the combined effect of the component of the cargo's weight along the inclined plane and the surface friction. The task breakdown module allocates the maximum response delay based on the maximum allowable passage time corresponding to each sub-task. The maximum response delay is set as the product of the maximum allowable passage time and a preset response reserve coefficient, where the response reserve coefficient is set to a value... The selection criterion is to retain a certain communication and execution buffer when the time limit is approaching. The maximum response latency represents the longest allowed response time from the moment the subtask starts executing the instruction to the completion of the subtask job, and this value is written in the task description body of the subtask.

[0028] When determining each subtask and its corresponding maximum allowable passage time, the task decomposition module also calculates the minimum travel speed requirement for each subtask based on the path segment length and the maximum allowable passage time. The path segment length is the Euclidean distance between the starting and ending spatial coordinates. The minimum travel speed requirement is obtained by dividing the path segment length by the maximum allowable passage time, in meters per second. The task decomposition module binds the minimum travel speed requirement as a dynamic constraint to the subtask. This binding process is achieved by writing the minimum travel speed requirement into the speed constraint field of the subtask's task description body. Subsequently, after the dynamic allocation module associates the main agent and auxiliary agent with the subtask, the minimum travel speed requirement bound in the subtask will be sent to the motion controllers of the main agent and auxiliary agent along with the task instructions, serving as the lower limit constraint for speed during motion control.

[0029] In specific implementation, please refer to Figure 3 The dynamic allocation module obtains the sequence of subtasks that have been decomposed from the task decomposition module and retrieves the registration information of all available agents from the heterogeneous agent set. The agents in the heterogeneous agent set differ in their capability attributes, which are reflected in differences in the upper limit of load capacity, maximum communication latency tolerance, and kinematic constraint parameters. The dynamic allocation module independently performs the selection process of the main agent and auxiliary agents for each subtask to be allocated.

[0030] The dynamic allocation module obtains the navigation distance between the current pose of each agent in the heterogeneous agent set and the starting point of the subtask. The current pose is obtained from the real-time data reported by the localization modules of each agent, including the agent's position coordinates and orientation angle in 3D space. The starting point of the subtask is the spatial coordinates of the starting point of the corresponding path segment of the subtask. The navigation distance is calculated using the A* path search algorithm based on a drivable grid map. The dynamic allocation module uses the agent's current position as the search starting point and the starting spatial coordinates of the subtask as the search target point, searches for an optimal path that avoids static obstacles on the drivable grid map, and calculates the total length of this optimal path as the navigation distance. If the agent's current position is completely blocked by obstacles, preventing the search for a path to the target point, the navigation distance of that agent is marked as infinite, and that agent does not participate in the subsequent calculation of the adaptation weight coefficient.

[0031] The dynamic allocation module obtains the ratio of each agent's load capacity to the required load capacity of the subtask. Load capacity, obtained from the agent's factory calibration parameters or real-time health monitoring, represents the maximum load capacity the agent can output under stable operating conditions, measured in Newtons. The required load capacity of the subtask is calculated by the task decomposition module and stored in the subtask's task description body. The dynamic allocation module divides the load capacity by the required load capacity to obtain a dimensionless ratio. When this ratio is greater than or equal to a preset load capacity threshold, it indicates that the agent possesses the basic condition to independently bear the weight of the goods in the subtask. The preset load capacity threshold is 1.0, meaning the load capacity is at least equal to the required load capacity.

[0032] The dynamic allocation module obtains the difference between the communication latency threshold for each agent and the maximum allowable response latency for each subtask. The communication latency threshold, obtained from the agent's communication module registration information, represents the maximum tolerable one-way latency, in milliseconds, for real-time communication between the agent and the system's central control node. The maximum allowable response latency for each subtask is allocated by the task decomposition module and stored in the subtask's task description body. The dynamic allocation module calculates the difference between the maximum response latency and the communication latency threshold. This difference reflects the communication redundancy latency margin available to the agent when executing the subtask; a larger difference indicates more lenient communication conditions.

[0033] The dynamic allocation module calculates the adaptation weight coefficient for each agent to the sub-task based on the ratio of navigation distance, load capacity to required carrying capacity, and the difference between the communication latency threshold and the maximum response latency. The formula for calculating the adaptation weight coefficient is as follows:

[0034]

[0035] in, Indicates the first The agent targets the first The adaptation weight coefficient for each subtask. Indicates the first The current pose of the first agent and the first The navigation distance between the starting points of each sub-task, in meters. This is a zero-offset constant for navigation distance, with a value of [value missing]. The meter is used to avoid the denominator becoming meaningless when the navigation distance is close to zero. Indicates the first The load capacity of the first agent and the first The ratio of the required capacity of each sub-task. Indicates the first The maximum allowable response latency for each subtask minus the first... The difference in communication delay thresholds between agents, in milliseconds. This is a time base constant, with a value of Milliseconds are used to normalize time difference-related terms to a dimensionless interval. , , These are the contribution coefficients for navigation distance, load capacity, and communication redundancy, respectively. The range of values ​​is to , The range of values ​​is to , The range of values ​​is to The values ​​of the three contribution coefficients satisfy the following: The normalized constraints. In the formula for calculating the adaptation weight coefficient, the first term measures the ease with which the agent can reach the starting point of the operation in the form of the reciprocal of the navigation distance; the second term measures the adaptation margin of the agent in terms of carrying capacity in the form of saturation growth; and the third term measures the adaptation redundancy of the agent in terms of communication latency.

[0036] After calculating the adaptation weight coefficients for all agents in the heterogeneous agent set for the current subtask, the dynamic allocation module sorts all adaptation weight coefficients in descending order. The agent with the highest adaptation weight coefficient is designated as the primary agent. The module then checks the communication protocol compatibility between the agent with the second-highest adaptation weight coefficient and the primary agent. This compatibility check is performed by querying the list of supported communication protocol stacks recorded in the agent's registration information. If the communication protocol stack list of the agent with the second-highest adaptation weight coefficient contains the identifier of the communication protocol used by the primary agent, the two are considered compatible, and the agent with the second-highest adaptation weight coefficient is designated as the auxiliary agent. If the communication protocol of the agent with the second-highest adaptation weight coefficient is incompatible with that of the primary agent, the module checks the next agent in the descending order of adaptation weight coefficients until an agent compatible with the primary agent's communication protocol is found and designated as the auxiliary agent. If, after traversing all remaining agents, no auxiliary agent compatible with the main agent's communication protocol is found, the dynamic allocation module marks the subtask as a single-agent execution mode, and the main agent completes the subtask independently. The dynamic allocation module associates and binds each subtask with its selected main agent identifier and auxiliary agent identifier, forming a collaborative agent group configuration record for that subtask, and synchronously distributes the collaborative agent group configuration record to the task scheduling module and the interaction perception module.

[0037] In practical implementation, the interactive sensing module initiates the sensing data acquisition process during the execution of sub-tasks by the collaborative intelligent agents. The interactive sensing module collects physical interaction force data through a six-dimensional force sensor installed at the end effector flange of each intelligent agent. The six-dimensional force sensor is installed between the end effector flange and the clamping mechanism of the intelligent agent, and can simultaneously measure force components and torque components in three orthogonal directions in three-dimensional space. The data sampling frequency of the six-dimensional force sensor is synchronized with the servo cycle of the intelligent agent's motion controller, outputting a set of six-dimensional force vectors in each servo cycle. The interactive sensing module performs gravity compensation processing on the six-dimensional force vectors. This gravity compensation process subtracts the gravity component from the original force data using pre-calibrated installation posture of the six-dimensional force sensor and the self-weight parameters of the clamping mechanism, obtaining the net interaction force data generated by physical contact between intelligent agents or contact with the cargo. The physical interaction force data includes the contact force between the main intelligent agent and the auxiliary intelligent agent, as well as the contact force between the intelligent agent and the cargo corresponding to the target handling task. Each contact force is decomposed into normal force components and tangential force components. The normal force component is the force component along the normal direction of the contact surface, and the tangential force component is the force component along the tangential plane of the contact surface and consistent with the sliding direction. The decomposition process is completed based on the calibration relationship matrix between the six-dimensional force sensor coordinate system and the normal of the contact surface at the end of the intelligent agent.

[0038] The interactive sensing module acquires the 3D spatial coordinates and orientation angles of each agent in the collaborative agent group in real time through an ultra-wideband positioning module or a visual odometry. The ultra-wideband positioning module consists of fixed anchor nodes deployed in the operating environment and tag nodes installed on each agent. The time-of-flight ranging values ​​between the tag nodes and multiple fixed anchor nodes are used to calculate the 3D spatial coordinates of the tag nodes using a trilateration algorithm. The visual odometry acquires environmental image features through a forward-facing stereo camera installed on the agent, and outputs the agent's 3D spatial coordinates and orientation angles in a local visual coordinate system using feature matching and pose estimation across consecutive frames. The 3D spatial coordinates are represented using the X, Y, and Z axes of a Cartesian coordinate system, and the orientation angles are represented using Euler angles: yaw, pitch, and roll. The raw positioning data acquired by the interactive sensing module each carries an independent local coordinate system, requiring a coordinate system transformation. This transformation is performed using a pre-calibrated homogeneous transformation matrix from each sensor's coordinate system to the world coordinate system, converting all the 3D spatial coordinates and orientation angles output by the ultra-wideband positioning module or visual odometry to the world coordinate system. After completing the coordinate system transformation, the interactive perception module calculates the relative distance and relative attitude angle between each pair of agents as relative pose data. For agents A and B, the relative distance is obtained by the Euclidean distance between the three-dimensional spatial coordinates of agent A and agent B in their world coordinate system, and the relative attitude angle is obtained by the difference between the orientation angles of agent A and agent B. The difference in attitude angles needs to be converted to the range of [-π, π].

[0039] The conflict prediction module obtains physical interaction force data and relative pose data from the interaction perception module to construct a conflict probability field for the collaborative intelligent agent group within a shared workspace. The shared workspace is defined as a three-dimensional spatial region containing the current motion range of all agents in the collaborative intelligent agent group, and this three-dimensional spatial region is discretized into cubic grid cells of equal size.

[0040] The conflict prediction module predicts the trajectory of each agent in the next control cycle based on its current velocity vector and relative pose data. The velocity vector is obtained in real time from the agent's motion controller and includes three components of the linear velocity vector and three components of the angular velocity vector. Starting from the agent's current 3D spatial coordinates, the conflict prediction module uses the direction of the linear velocity vector as the direction of motion, and the product of the magnitude of the linear velocity vector and the control cycle duration as the step size. It then calculates the predicted position point at the end of the next control cycle along the direction of motion and uses the straight line segment connecting the current 3D spatial coordinates to the predicted position point as the centerline of the predicted trajectory.

[0041] The conflict prediction module corrects the width expansion coefficient of the predicted motion trajectory based on the normal and tangential force components in the physical interaction force data. The initial value of the width expansion coefficient is set to the radius of the agent's circumscribed cylinder plus a preset basic safety margin value, which is 0.05 meters. The width expansion coefficient increases linearly with the absolute value of the tangential force component, and the slope of this linear increase is the width expansion sensitivity coefficient, which is set to 0.002 meters per Newton. The upper limit of the width expansion coefficient increase does not exceed half of the minimum safe obstacle avoidance distance between the primary and secondary agents. The minimum safe obstacle avoidance distance between the primary and secondary agents is read from the system safety parameter library, which stores a table of safe distances at different speed ranges calculated based on the agents' maximum braking capacity. The conflict prediction module interpolates the minimum safe obstacle avoidance distance from the table based on the current speeds of the two agents. When the normal force component shows a stretching trend and exceeds a preset stretching threshold, the width expansion coefficient remains unchanged and is not adjusted with changes in the tangential force component. The preset stretching threshold is 10 Newtons.

[0042] The conflict prediction module uses the centerline of each agent's predicted trajectory as the field source and the width expansion coefficient as the field radius to generate a single-agent potential field for each agent within the shared workspace. The potential field value at the centerline is set to its maximum value of 1.0, and decreases radially outward from the centerline with increasing distance. A Gaussian kernel function is used as the decay function, with its standard deviation parameter set to one-third of the width expansion coefficient. When the radial distance exceeds three times the width expansion coefficient, the potential field value is truncated to zero. The conflict prediction module then superimposes the potential field values ​​of all single-agent potential fields generated by all agents at the same grid cell using arithmetic summation to obtain the superimposed conflict probability field.

[0043] The conflict prediction module marks grid cell regions where the superimposed potential field value exceeds a second preset threshold as high-probability regions in the conflict probability field. The second preset threshold is dynamically determined based on the potential field value distribution of all grid cells within the shared workspace, and is set to 1.5 times the median potential field value of all grid cells.

[0044] The eigenvalues ​​of the conflict probability field are used to quantitatively describe the overall intensity level of the field. The eigenvalues ​​are calculated as the weighted root mean square of the potential field values ​​at all grid points within the conflict probability field, using the following formula:

[0045]

[0046] in, The field strength eigenvalue represents the collision probability field, and its unit is a dimensionless quantity. The total number of grid cells covering the conflict probability area; Indicates the first The potential field value after superposition of individual grid cells; Indicates the first The weight coefficient of each grid cell, the weight coefficient The weighting coefficient is determined by the reciprocal of the shortest spatial distance from the center point of each grid cell to the center line of the predicted motion trajectory of each agent. Grid cells with a distance less than the width expansion coefficient have a weighting coefficient of 1.0, while those with a distance exceeding three times the width expansion coefficient have a weighting coefficient of 0.1. The intermediate segments are determined by linear interpolation. The conflict prediction module calculates the field strength characteristic value once per control cycle. The calculation results, along with the high-probability region distribution information of the conflict probability field, are sent to the collaborative adjustment module.

[0047] In practical implementation, the collaborative adjustment module receives the field strength characteristic values ​​of the conflict probability field and the distribution information of high-probability regions in the conflict probability field from the conflict prediction module. The collaborative adjustment module extracts connected components from all high-probability regions in the conflict probability field. High-probability regions are composed of grid cells marked by the conflict prediction module as having a potential field value higher than a second preset threshold after superposition. Connected component extraction employs a six-connectivity region growth algorithm based on a 3D mesh. The collaborative adjustment module traverses all grid cells marked as high-probability regions, grouping adjacent grid cells belonging to the same connected component. Adjacency is determined by grid cells sharing faces, edges, or vertices in the 3D mesh. Each connected component consists of a set of spatially connected grid cells.

[0048] The collaborative adjustment module calculates the average potential field value within each extracted connected component and uses this average potential field value as the field strength characteristic value of that connected component. The formula for calculating the average potential field value is:

[0049]

[0050] in, Indicates the first The field strength eigenvalues ​​of a connected domain, in dimensionless quantities; Indicates the first The total number of grid cells contained within each connected domain; Indicates the first The set of indices of all grid cells in a connected domain; Indicates the first The potential field value after superposition of individual grid cells is calculated and output by the conflict prediction module when constructing the conflict probability field. The collaborative adjustment module calculates a field strength eigenvalue for each connected component. The collaborative adjustment module compares the field strength characteristic value of each connected component with a first preset threshold. This first preset threshold is stored in the system security configuration parameter table and is set to 0.65. This value is based on the statistical analysis of historical collaborative handling operation data, which shows that when the average potential field value of the conflict probability field exceeds 0.65, the probability of actual physical interference increases significantly. When the field strength characteristic value of any connected component exceeds the first preset threshold, the collaborative adjustment module identifies the conflicting agent combination corresponding to that connected component. The identification process for the conflicting agent combination is as follows: the collaborative adjustment module obtains the contribution distribution of all grid cells within the connected component in the potential field of each individual agent. Agents whose potential field value as a percentage of the summed potential field value of the grid cells exceeds a preset contribution ratio threshold are included in the conflicting agent combination. The preset contribution ratio threshold is set to 0.25. By summarizing the agent contributors corresponding to all grid cells within the connected component and removing duplicate identifiers, a conflicting agent combination composed of multiple agent identifiers is obtained.

[0051] The collaborative adjustment module obtains the priority of the subtasks currently being executed by each agent in the conflicting agent group. The priority of subtasks is specified by the task decomposition module based on execution order dependencies and time urgency when generating subtasks, with priorities categorized into high, medium, and low levels. The collaborative adjustment module queries the task scheduler for the subtask identifier currently bound to each agent and its corresponding priority level. The module compares the subtask priorities of each agent in the conflicting agent group, temporarily extending the maximum allowed passage time for lower-priority subtasks and compressing the maximum allowed passage time for higher-priority subtasks. The temporary extension of the maximum allowed passage time is achieved by modifying the time constraint parameters of the corresponding subtasks in the task scheduler. The extension amount is set to the absolute value of the compression amount of the higher-priority subtask, with the total extension and compression amount remaining zero. The compression amount is calculated as 10% of the original maximum allowed passage time of the subtask. When priorities are the same, the objects to be extended and compressed are determined based on the lexicographical order of the agent identifiers.

[0052] After temporarily extending and compressing the maximum allowed passage time, the collaborative adjustment module notifies the task scheduler to recalculate the execution sequence of subsequent unexecuted subtasks based on the updated maximum allowed passage time, and synchronously updates the conflict probability field. The synchronously updated conflict probability field is generated by the conflict prediction module recalculating the potential field of a single agent based on the new subtask execution distribution and the predicted motion trajectory of the agents, and then superimposing these values. The collaborative adjustment module obtains the field strength feature value of the corresponding connected component in the updated conflict probability field and compares it again with the first preset threshold. If the updated field strength feature value still exceeds the first preset threshold, the collaborative adjustment module further triggers a local replanning of the motion trajectory of the conflicting agent combination. Local replanning is executed by the trajectory planner, invoked by the collaborative adjustment module. The trajectory planner uses the current 3D spatial coordinates of each agent in the conflicting agent group as its starting point and the endpoint spatial coordinates of each subtask as its temporary target point. Within the shared workspace, it generates a new path based on the elastic band trajectory planning method. During generation, a repulsive potential cost is introduced, based on the predicted trajectories of the other agents. This cost source is the high-probability region of the conflict probability field. The final output is a collision-free trajectory that reduces the field strength characteristic value of the conflict probability field below a first preset threshold. The trajectory generated by local replanning replaces the original planned trajectory and is then sent to the motion controllers of each agent in the conflicting agent group for execution.

[0053] In practical implementation, the collaborative obstacle avoidance control module is activated and executed after insertion during the collaborative safety adjustment phase. The module acquires the contour point cloud and velocity vectors of dynamic obstacles in front of the main agent. The contour point cloud is acquired using a LiDAR sensor mounted forward of the main agent. The LiDAR sensor emits multi-line laser beams and receives reflected echoes, generating 3D point cloud data of the environment. The module performs ground point filtering on the raw 3D point cloud data, separating ground points from non-ground points using a random sampling consensus algorithm, retaining non-ground points as candidate obstacle points. Euclidean clustering is then performed on the candidate obstacle points, with a clustering distance threshold set to 0.15 meters, clustering spatially close point clouds into independent obstacle targets. For each clustered obstacle target, the module calculates its bounding box, represented using an axis-aligned bounding box. The contour point cloud of the bounding box is a set of coordinates of the eight vertices of the bounding box and sampled points on its surface. The velocity vector is obtained by the main agent calculating the change in the centroid position of the bounding box of the same obstacle target between two consecutive lidar scan frames. The magnitude of the velocity vector is the distance of the centroid position change divided by the time interval between the two frames, and the direction of the velocity vector is the normalized result of the direction vector of the centroid position change.

[0054] The collaborative obstacle avoidance control module calculates the first remaining time for the main agent to collide with the dynamic obstacle at its current speed based on the contour point cloud and velocity vector. The module extracts the main agent's current velocity vector and unifies its 3D spatial coordinates and orientation angle with the dynamic obstacle's contour point cloud and velocity vector in the same world coordinate system. Employing a velocity-obstacle-based collision time calculation method, the module simplifies the main agent's shape into a spherical envelope with its geometric center as the center and the radius of its circumscribed sphere as the radius. The dynamic obstacle's contour point cloud is converted into a convex polyhedral envelope. The difference between the main agent's current velocity vector and the dynamic obstacle's velocity vector constitutes the relative velocity vector. Starting from the center point of the main agent's spherical envelope, a ray is emitted along the direction of the relative velocity vector. The intersection point of the ray and the dynamic obstacle's convex polyhedral envelope is calculated. The distance from the starting point to the intersection point is divided by the magnitude of the relative velocity vector to obtain the first remaining time for the main agent to collide with the dynamic obstacle at its current speed. If the ray has no intersection with the envelope of the convex polyhedron, then the first remaining time is set to infinity.

[0055] When the remaining time is less than the safety time threshold, the cooperative obstacle avoidance control module generates a lateral avoidance offset vector. The safety time threshold is determined by the sum of the shortest braking time required for the agent to come to a complete stop from its current speed and the communication delay. The shortest braking time is calculated using the agent's maximum braking deceleration rate parameter and the current speed, with the maximum braking deceleration rate parameter set to 2.5 meters per square second. The direction of the lateral avoidance offset vector is perpendicular to the yaw direction of the main agent's current orientation angle and points to the opposite side of the dynamic obstacle's movement direction. The magnitude of the lateral avoidance offset vector is calculated by referring to the sum of the half-width of the dynamic obstacle bounding box in the yaw direction, the radius of the main agent's circumscribed sphere, and a preset lateral safety additional distance, set to 0.3 meters. The cooperative obstacle avoidance control module superimposes the generated lateral avoidance offset vector into the main agent's current trajectory planning and simultaneously broadcasts the lateral avoidance offset vector to the auxiliary agent via a low-latency communication link. The broadcast message includes the three component values ​​of the lateral avoidance offset vector and a timestamp indicating when the superposition takes effect.

[0056] The cooperative obstacle avoidance control module controls the auxiliary agent to synchronously superimpose a lateral avoidance offset vector while maintaining a constant relative attitude angle with the master agent, forming an overall formation offset motion. The relative attitude angle is the difference between the orientation angle of the auxiliary agent and the orientation angle of the master agent in the yaw direction. This difference is determined and stored in the formation configuration parameters when the master and auxiliary agents form a cooperative formation. After receiving the lateral avoidance offset vector, the cooperative obstacle avoidance control module converts the lateral avoidance offset vector into an offset displacement command executable by the auxiliary agent's motion controller. In the next control cycle, the auxiliary agent superimposes the offset displacement command onto the original trajectory tracking command and executes it, while maintaining a constant relative attitude angle with the master agent. The auxiliary agent adjusts the yaw angle through its own attitude controller in a closed loop to ensure that it always follows the yaw angle of the master agent, superimposed with the yaw angle offset value set in the formation configuration parameters. After the main agent and the auxiliary agent complete the synchronous offset, the collaborative agent group as a whole avoids the dynamic obstacle by translation. After the dynamic obstacle is passed, the collaborative obstacle avoidance control module generates a reverse offset vector in the next control cycle to restore the original motion trajectory.

[0057] During the execution of sub-tasks by the collaborative intelligent agents, the collaborative load balancing module monitors in real time the first current value of the master agent's drive motor and the second current value of the auxiliary agent's drive motor. The drive motor current value is acquired through a current sensor built into the motor driver. The current sensor outputs the instantaneous phase current value of the drive motor in each current sampling period. The collaborative load balancing module takes the root mean square value of the three-phase current as the drive motor current value. The first current value of the master agent's drive motor reflects the actual load torque level of the master agent during the current sub-task execution, while the second current value of the auxiliary agent's drive motor reflects the actual load torque level of the auxiliary agent.

[0058] The collaborative load balancing module calculates the instantaneous deviation between the first current value and the second current value. The formula for calculating the instantaneous deviation is:

[0059]

[0060] in, This represents the instantaneous deviation of the drive motor current between the main intelligent agent and the auxiliary intelligent agent, expressed in amperes. The first current value of the driving motor of the main intelligent agent is represented in amperes. It is collected in real time by the current sensor of the main intelligent agent motor driver and output after being processed by a low-pass filter with a cutoff frequency of 50 Hz. The second current value, in amperes, represents the drive motor's current value for the auxiliary agent. It is acquired in real-time by the current sensor of the auxiliary agent's motor driver and output after being processed by a low-pass filter with a cutoff frequency of 50 Hz. The collaborative load balancing module determines whether the instantaneous deviation exceeds the load balancing threshold. The load balancing threshold is set according to the maximum allowable load imbalance during collaboration. The load balancing threshold stored in the parameter configuration file of the collaborative load balancing module is 15% of the rated current values ​​of the two agent drive motors. When the rated current value of either agent is unavailable, the load balancing threshold takes the alternative fixed value of 5.0 amperes.

[0061] When the instantaneous deviation exceeds the load balancing threshold, the collaborative load balancing module generates a torque redistribution command. The core of this command is to issue a torque adjustment to the auxiliary agent. This adjustment is calculated by multiplying the instantaneous deviation by a proportional adjustment coefficient, which is the reciprocal of 0.8 amperes, ensuring the torque adjustment is proportional to the current deviation. The collaborative load balancing module sends the torque redistribution command to the auxiliary agent's motor controller via a low-latency communication link. This low-latency link is a dedicated point-to-point high-speed data link within the collaborative agent group, using Gigabit Ethernet or wireless fidelity direct connection, with a communication latency of no more than 1 millisecond. Upon receiving the torque redistribution command, the auxiliary agent's motor controller adjusts the output torque of its drive motor according to the torque adjustment. The increase or decrease in output torque is calculated by multiplying the torque adjustment by the torque constant of the auxiliary agent's drive motor. The torque constant is an inherent parameter of the auxiliary agent's drive motor, obtained from the motor nameplate or factory calibration data. The motor controller of the auxiliary intelligent agent immediately applies the new output torque setpoint at the beginning of the next control cycle, so that the first current value and the second current value tend to be consistent in the next control cycle. The collaborative load balancing module repeatedly monitors and compares the first current value and the second current value in each control cycle, forming a closed-loop regulation process.

[0062] Upon detecting a communication interruption, the local autonomous collaboration module executes the communication interruption response procedure. The local autonomous collaboration module maintains a heartbeat signal counter. The primary agent checks whether it has received a heartbeat signal broadcast by the secondary agent within each control cycle. The heartbeat signal is a short data packet sent by the secondary agent at a fixed heartbeat interval, set to 50 milliseconds. The heartbeat signal contains the secondary agent's identifier and the latest timestamp. When the primary agent fails to receive a heartbeat signal from the secondary agent for several consecutive control cycles, it determines that the communication link with the secondary agent is interrupted. The number of consecutive control cycles required to determine a communication interruption is set based on the relationship between the heartbeat interval and the control cycle duration. Using a heartbeat interval of 50 milliseconds and a control cycle of 10 milliseconds as a baseline, an interruption determination is triggered when five consecutive control cycles of lost heartbeat signals occur, with the cumulative time exceeding 50 milliseconds.

[0063] Upon determining that the communication link has been interrupted, the primary agent switches to local predictive control mode. In local predictive control mode, the primary agent retrieves the motion state parameters of the most recently received auxiliary agent from its local cache. These parameters include the auxiliary agent's three-dimensional spatial coordinates, velocity vector, and orientation angle, and are carried in the last successfully received heartbeat signal before the communication interruption. Based on these motion state parameters, the primary agent calculates the auxiliary agent's predicted trajectory during the interruption. The calculation method involves using the last known three-dimensional spatial coordinates of the auxiliary agent as the starting point and the last known velocity vector as a constant velocity input, extrapolating step-by-step according to the control cycle time step to obtain the predicted position point of the auxiliary agent at each time step. These predicted position points are then connected in chronological order to form the predicted trajectory. The calculation of the predicted trajectory continues until communication is restored or a preset maximum calculation duration is reached (set to 5 seconds).

[0064] The primary agent reduces its movement speed and increases the safety clearance between itself and the predicted trajectory based on the proximity between its actual trajectory and the predicted trajectory. The proximity is measured using the Euclidean distance from the primary agent's current 3D spatial coordinates to the nearest point on the predicted trajectory. The reduction rate of the primary agent's movement speed is determined according to the proximity level: when the Euclidean distance is greater than 2.0 meters, the movement speed remains unchanged; when the Euclidean distance is between 1.0 and 2.0 meters, the movement speed is 70% of the original speed setting; when the Euclidean distance is between 0.5 and 1.0 meters, the movement speed is 40% of the original speed setting; and when the Euclidean distance is less than 0.5 meters, the movement speed is 10% of the original speed setting, triggering a deceleration and stop preparation. The safety clearance is the minimum distance that should be maintained between the primary agent and the auxiliary agent and the predicted trajectory. The initial value of the safety clearance is the default safety clearance value, which is 1.5 meters. This value is based on twice the minimum reaction distance required for the primary and auxiliary agents to achieve emergency obstacle avoidance using their own sensors without communication. When the proximity changes, the safety protection gap is adjusted by adding a dynamic offset to the default value. The dynamic offset is proportional to the reciprocal of the proximity and the proportionality coefficient is 0.5 square meters.

[0065] When the auxiliary agent re-establishes the communication connection, the primary agent and the auxiliary agent exchange trajectory logs from the interruption period. The trajectory logs contain timestamps for each control cycle during the interruption, the auxiliary agent's 3D spatial coordinates, and movement velocity vector. These logs are continuously recorded locally by the auxiliary agent and packaged and sent to the primary agent after communication is restored. The primary agent also sends its own trajectory logs from the interruption period to the auxiliary agent. The local autonomous collaboration module dynamically updates the default value of the safety protection gap based on the maximum approach distance in the trajectory logs. The maximum approach distance refers to the shortest Euclidean distance between the actual movement trajectory of the primary agent and the actual movement trajectory of the auxiliary agent at all times during the interruption. The default value update strategy for the safety protection gap is to multiply the maximum approach distance by an update coefficient of 1.2 to obtain the new default safety protection gap value, ensuring that the updated default safety protection gap value is slightly larger than the historical closest distance, thus preserving safety redundancy. If the maximum approach distance is less than the original default safety protection gap value, the original default safety protection gap value remains unchanged.

[0066] After completing the target transportation task, the collaboration efficiency evaluation module initiates the post-task evaluation process. The module records the total time elapsed from acquiring the target transportation task to the completion of the last subtask. The recording method is to use the timestamp of the target transportation task received by the task parsing module as the start time and the timestamp of the completion of the last subtask reported by the task scheduler as the end time. The total time is calculated by the time difference between the end time and the start time, in seconds. The module also records the number of conflicts and the average duration of conflicts that occurred during the entire operation. The number of conflicts is obtained by counting the total number of times the collaborative adjustment module triggered the collaborative safety adjustment phase during the operation. The average duration of conflicts is calculated by taking the arithmetic mean of all durations from the start of each collaborative safety adjustment phase until the field strength characteristic value of the conflict probability field falls below the first preset threshold, in seconds.

[0067] The collaboration efficiency evaluation module calculates the overall collaboration efficiency coefficient for this task based on the total time, number of conflicts, and average conflict duration. The overall collaboration efficiency coefficient is calculated using the total time, number of conflicts, and average conflict duration as input variables, mapped to a dimensionless coefficient between zero and one through a weighted scoring function. In the weighted scoring function, the score for the total time dimension is obtained by dividing the theoretical minimum completion time specified in the task parameters of the target transport task by the actual total time. The score for the number of conflicts dimension is calculated using an exponential decay function, where the base of the decay function is the natural constant, and the exponent is the number of conflicts multiplied by the decay coefficient, with a value of 0.1. The score for the average conflict duration dimension is calculated by subtracting the actual average conflict duration from the baseline duration and then dividing by the normalized baseline duration, where the baseline duration is 20% of the average time spent executing a single subtask. The weighted coefficients for the three dimensions are 0.5, 0.3, and 0.2, respectively. The closer the overall collaboration efficiency coefficient is to 1.0, the higher the collaboration efficiency of this task.

[0068] The collaborative efficiency evaluation module associates the overall collaborative efficiency coefficient with the task parameters of the target handling task and stores them in the historical task database. Task parameters include cargo weight, spatial constraint feature vectors of the handling path, and temporal constraint feature vectors. The spatial constraint feature vectors consist of the handling path length, standard deviation of the path slope, average passable width, and static obstacle density. The temporal constraint vectors consist of the total time limit requirement, time window type, and time urgency index. Each record in the historical task database contains a task parameter field, a collaborative agent allocation scheme field, and an overall collaborative efficiency coefficient field.

[0069] When a new transport task with a similarity higher than the similarity threshold is received, the collaborative efficiency evaluation module retrieves the allocation strategy with the highest overall collaborative efficiency coefficient from the historical task database as a reference for the initial allocation scheme. The calculation of task parameter similarity uses the cosine similarity method. Cosine similarity is calculated for both spatial constraint feature vectors and temporal constraint vectors in the task parameters, and then a weighted average is taken. The weight of the spatial constraint feature vector is 0.6, and the weight of the temporal constraint vector is 0.4. The similarity threshold is set to 0.75. When multiple historical records with a similarity higher than the similarity threshold are retrieved, they are sorted in descending order of overall collaborative efficiency coefficient. The collaborative agent allocation scheme in the historical record with the highest overall collaborative efficiency coefficient is used as a reference for the initial allocation scheme, serving as an initial seed for the dynamic allocation module during the calculation of the adaptation weight coefficient.

[0070] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A system for dynamic task allocation and collaboration based on intelligent agents, characterized in that, include: The task parsing module obtains the task parameters of the target transportation task and parses the spatial and temporal constraints in the target transportation task. The task decomposition module decomposes the target transportation task into multiple sub-tasks with execution order dependencies based on the spatial constraints and the time constraints. The dynamic allocation module dynamically selects the main and auxiliary agents to execute each sub-task from the heterogeneous agent set based on the current pose, load capacity and communication latency threshold of each agent, forming a cooperative agent group for the target transportation task. The interaction perception module collects physical interaction force data and relative pose data between the agents in real time during the execution of the sub-task by the collaborative intelligent agent group. The conflict prediction module constructs a conflict probability field of the cooperative intelligent agent group in the shared work space based on the physical interaction force data and the relative pose data. The collaborative adjustment module, when the field strength characteristic value of the conflict probability field exceeds the first preset threshold, inserts a collaborative safety adjustment phase into the subsequent unexecuted sub-tasks to adjust the execution timing or movement trajectory of the main intelligent agent and the auxiliary intelligent agent.

2. The agent-based task dynamic allocation and cooperation system according to claim 1, characterized in that, Based on the spatial and temporal constraints, the target transportation task is decomposed into multiple sub-tasks with execution order dependencies, specifically including: Based on the spatial constraints, the transport path corresponding to the target transport task is discretized into multiple path segments, and the starting spatial coordinates and ending spatial coordinates are marked for each path segment. Based on the time constraints, assign a maximum allowed passage time to each path segment; Each path segment and its corresponding maximum allowed passage time are determined as an independent subtask, and the execution order dependency between multiple subtasks is determined based on the adjacency relationship between the starting point spatial coordinates and the ending point spatial coordinates. Calculate the required load-bearing capacity for each sub-task based on the cargo weight, slope, and friction coefficient of each path segment. Based on the maximum allowed throughput time corresponding to each subtask, a maximum response delay is assigned to the subtask, where the maximum response delay represents the longest allowed response time from the start of execution of the subtask to its completion.

3. The agent-based task dynamic allocation and cooperation system according to claim 2, characterized in that, When determining each path segment and its corresponding maximum allowed passage time as an independent subtask, the minimum travel speed requirement for each subtask is calculated based on the length of the path segment and the maximum allowed passage time, and is then bound to the main agent and the auxiliary agent as a dynamic constraint condition for that subtask.

4. The agent-based dynamic task allocation and cooperation system according to claim 1, characterized in that, The process of dynamically selecting the main and auxiliary agents to execute each subtask from a heterogeneous set of agents, based on each agent's current pose, load capacity, and communication latency threshold, specifically includes: Obtain the navigation distance between the current pose of each agent in the heterogeneous agent set and the starting point of the subtask; Based on the ratio of the navigation distance, the load capacity, and the required carrying capacity of the subtask, and the difference between the communication delay threshold and the maximum allowable response delay of the subtask, calculate the adaptation weight coefficient of each agent for the subtask. The agent with the highest adaptation weight coefficient is determined as the main agent, and the agent with the second highest adaptation weight coefficient and whose communication protocol is compatible with the main agent is determined as the auxiliary agent.

5. A system for dynamic allocation and collaboration of tasks based on intelligent agents according to claim 1, characterized in that, During the execution of the sub-task by the collaborative intelligent agent group, physical interaction force data and relative pose data between the agents are collected in real time, specifically including: By installing a six-dimensional force sensor at the end of each intelligent agent, the physical interaction force data between the main intelligent agent and the auxiliary intelligent agent, as well as between the intelligent agent and the goods corresponding to the target handling task, are collected. The three-dimensional spatial coordinates and orientation angles of each agent in the collaborative intelligent agent group are obtained in real time through an ultra-wideband positioning module or a visual odometry. After performing a coordinate system transformation on the three-dimensional spatial coordinates and the orientation angle, the relative distance and relative attitude angle between each pair of intelligent agents are calculated as the relative pose data.

6. The agent-based task dynamic allocation and cooperation system according to claim 1, characterized in that, Based on the physical interaction force data and the relative pose data, a conflict probability field is constructed for the cooperative intelligent agent group within the shared workspace, specifically including: Based on the current movement velocity vector of each agent and the relative pose data, predict the predicted motion trajectory of the agent in the next control cycle; The width expansion coefficient of the predicted motion trajectory is corrected based on the normal force component and tangential force component in the physical interaction force data. Using the centerline of the predicted motion trajectory of each agent as the field source and the width expansion coefficient as the field radius, a single agent potential field is generated for each agent in the shared work space. The potential fields of all agents are superimposed, and the regions where the superimposed potential field value is higher than the second preset threshold are marked as high probability regions of the conflict probability field.

7. A system for dynamic allocation and collaboration of tasks based on intelligent agents according to claim 6, characterized in that, The width expansion coefficient increases linearly according to the magnitude of the tangential force component, and the upper limit of the increase does not exceed half of the minimum safe obstacle avoidance distance between the main intelligent agent and the auxiliary intelligent agent.

8. A system for dynamic allocation and collaboration of tasks based on intelligent agents according to claim 6, characterized in that, When the field strength characteristic value of the conflict probability field exceeds a first preset threshold, a collaborative safety adjustment phase is inserted into the subsequent unexecuted subtasks to adjust the execution timing or trajectory of the main agent and the auxiliary agent, specifically including: Extract the connected regions of all high-probability regions in the conflict probability field, and calculate the average potential field value in each connected region as the field strength feature value; When the field strength feature value of any connected domain exceeds the first preset threshold, the conflicting agent combination corresponding to the connected domain is identified. Based on the priority of the subtasks currently being executed by each agent in the conflict agent group, the maximum allowed passage time corresponding to the lower priority subtasks is temporarily extended, and the maximum allowed passage time corresponding to the higher priority subtasks is correspondingly compressed, thereby adjusting the execution sequence.

9. A system for dynamic allocation and collaboration of tasks based on intelligent agents according to claim 8, characterized in that, After temporarily extending the maximum allowed passage time, the conflict probability field is updated synchronously. If the updated field strength feature value still exceeds the first preset threshold, then the motion trajectory of the conflicting agents is further replanned locally.

10. A system for dynamic allocation and collaboration of tasks based on intelligent agents according to claim 1, characterized in that, The field strength characteristic value of the conflict probability field is the weighted root mean square value of the potential field values ​​of all grid points within the conflict probability field.

Images