Machine learning based multi-tier shuttle robot dynamic task allocation method

Abstract

The present application relates to the technical field of task allocation management of logistics storage, and more particularly to a multi-layer shuttle robot dynamic task allocation method based on machine learning, which comprises: calculating an ideal execution time for each combination of a to-be-assigned task and an idle robot; predicting a future motion trajectory of the robot, superimposing time-space occupation potential fields generated by all robots on a key resource point passed through by the future motion trajectory to obtain a time-space aggregation density of the key resource point; calculating a scheduling stability penalty term of the idle robot; adding the ideal execution time of the idle robot and the scheduling stability penalty in time to obtain a final evaluation value; and selecting an idle robot with the minimum final evaluation value to execute the to-be-assigned task, thereby effectively improving the accuracy of robot task allocation and the efficiency of robot task execution.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of task allocation management of logistics warehousing. More particularly, the present application relates to a dynamic task allocation method for multi-layer shuttle robots based on machine learning. BACKGROUND

[0002] A multi-layer shuttle robot system is a core device in a modern automated stereoscopic warehouse. It moves horizontally through shuttle cars deployed in the tracks of each layer of shelves and uses an elevator to achieve cross-layer vertical transportation, thereby completing high-density goods storage and retrieval operations. The overall operation efficiency of the system depends largely on the merits of its task allocation strategy, i.e., how to efficiently and reasonably allocate the storage and retrieval tasks to the idle shuttle robots in the system.

[0003] In the prior art, a common task allocation method is to use a contract net protocol (CNP) based on cost bidding. Under this protocol, the warehouse control system (WCS) broadcasts tasks to all idle robots, and the robots bid according to the physical distance from their own position to the target point of the task as the cost, and the WCS finally selects the robot with the lowest cost to execute the task.

[0004] However, when evaluating the cost, this method only considers the horizontal movement distance of the robot, but completely ignores the vertical traffic bottleneck caused by the elevator in the system. This leads to a short-sighted decision, i.e., a robot on a different floor, although its horizontal projection distance to the task point is very close, its real time cost to complete the task includes the waiting and riding time of the elevator, which may be higher. Therefore, this simplified cost model cannot reflect the complexity of cross-layer operations, often leading to robot congestion at the elevator entrance, long task waiting time, and thus seriously affecting the operation efficiency and throughput of the entire warehouse system.

[0005] Therefore, how to effectively improve the dynamic task allocation efficiency of multi-layer shuttle robots is a problem to be solved at present. SUMMARY

[0006] To solve the above technical problem of how to effectively improve the dynamic task allocation efficiency of multi-layer shuttle robots, the present application proposes a dynamic task allocation method for multi-layer shuttle robots based on machine learning, which comprises the following steps:

[0007] acquire task information to be allocated, state information of each elevator and state information of all idle robots, wherein the state information of the elevator comprises the number of queued robots at the entrance of the elevator;for each combination of the task to be allocated and the idle robot, calculate an ideal execution time, which comprises horizontal movement time between the idle robot and the task point position, and dynamic waiting time required for cross-layer operation predicted based on the number of queued robots and historical service data of the elevator;predict a future motion trajectory of the robot, superimpose the space-time occupation potential field generated by all robots on the key resource points passed through by the future motion trajectory to obtain the space-time aggregation density of the key resource points;calculate a scheduling stability penalty term of the idle robot, which is positively correlated with the space-time aggregation density of each key resource point on the future motion trajectory of the idle robot;add the ideal execution time of the idle robot and the scheduling stability penalty in time to obtain a final evaluation value;select the idle robot with the minimum final evaluation value to execute the task to be allocated.

[0008] The application provides a globally optimized dynamic task allocation method, which can effectively improve the task allocation accuracy of multi-layer shuttle robots and improve the operation efficiency. In the task allocation process, the method can accurately calculate the benchmark task time of the robot by comprehensively calculating the ideal execution time comprising the horizontal movement time and the dynamic waiting time predicted based on the number of queued elevators. On this basis, the application further introduces a scheduling stability penalty term based on the prediction of the future motion trajectory to evaluate potential traffic congestion, thereby constructing a more comprehensive task cost evaluation model. This method not only considers the execution efficiency of a single robot, but also takes into account the running harmony of the entire robot system, which can fundamentally avoid the elevator or key path congestion problem caused by improper decision-making, realizes the global optimization of robot task allocation, and significantly improves the operation efficiency and throughput of the entire warehouse system.

[0009] According to the multi-layer shuttle robot dynamic task allocation method based on machine learning provided by the application, the acquisition of the task information to be allocated, the state information of each elevator and the state information of all idle robots further comprises: collecting and preprocessing the running data of the elevator and the robot to obtain the state information of the elevator and the state information of the robot;wherein the state information of the elevator further comprises the single service time in the historical service data of the elevator;the state information of the robot comprises the current position, the running state, the running speed and the actual power;the task information to be allocated at least comprises the task point position, and the robot with an idle running state is taken as an idle robot.

[0010] According to the method for dynamically assigning tasks of multi-layer shuttle robots based on machine learning provided by the application, the horizontal movement time of the idle robot is obtained in the following manner: the Manhattan distance between the idle robot and the task point position is obtained, and the horizontal movement time between the idle robot and the task point position is obtained by taking the ratio of the Manhattan distance and the running speed of the idle robot.

[0011] The application determines the horizontal movement time between the idle robot and the task point position by Manhattan distance, which is highly consistent with the actual movement mode of the robot moving at right angles along the aisle of the shelf in the warehouse environment, so that the final time consumption result is closer to the real time consumption, and the accuracy of the ideal execution time calculation is effectively improved.

[0012] According to the method for dynamically assigning tasks of multi-layer shuttle robots based on machine learning provided by the application, the dynamic waiting time required for cross-layer operation is obtained based on the number of robots in the queue and the historical service data of the lifting machine, and the dynamic waiting time required for cross-layer operation is obtained by weighted summation of the product of the predicted waiting time of the last period and the actual waiting time of the current period, wherein the actual waiting time is the product of the number of robots in the queue of the lifting machine and the average single service time in the historical service data.

[0013] The application provides an accurate calculation method for the dynamic waiting time required for cross-layer operation, which can smooth transient fluctuations by weighted summation of the predicted value of the last period and the actual value of the current period, dynamically learn and adapt to the change trend of the service efficiency of the lifting machine, and improve the accuracy of the predicted time length.

[0014] According to the method for dynamically assigning tasks of multi-layer shuttle robots based on machine learning provided by the application, the ideal execution time is calculated, including:

[0015] ;

[0016] 、 respectively, the ideal execution time of the i-th idle robot for executing the to-be-assigned task j, the dynamic waiting time required for cross-layer operation, is the horizontal movement time between the current position of the i-th idle robot and the task point position for executing the to-be-assigned task j, is the full power of the idle robot, is the actual power of the i-th idle robot.

[0017] The application provides a precise ideal execution time calculation method, which introduces electric quantity as a penalty factor, increases the time cost of a robot with low electric quantity, intelligently reduces the probability of the robot being selected to execute a task, and avoids the risk of system blockage caused by the robot stopping due to low electric quantity during the task, thereby effectively prolonging the endurance of the robot with low electric quantity.

[0018] According to the multi-layer shuttle robot dynamic task allocation method based on machine learning provided by the application, the future motion trajectory of all idle robots is predicted according to a path planning algorithm.

[0019] According to the multi-layer shuttle robot dynamic task allocation method based on machine learning provided by the application, the acquisition method of the space-time occupation potential field generated by the robot comprises: acquiring the predicted position of the robot at each moment in the future motion trajectory; constructing a space-time occupation potential field model for representing the influence degree of the robot on a space point at each moment in the future motion trajectory, to obtain the space-time occupation potential field generated by the robot on the space point; wherein the space-time occupation potential field model is a Gaussian potential field model, and the space-time occupation potential field and the square of the distance from the space point to each predicted position of the robot are in an inverse exponential relationship.

[0020] According to the multi-layer shuttle robot dynamic task allocation method based on machine learning provided by the application, the calculation of the scheduling stability penalty term of the idle robot comprises:

[0021] ;

[0022] the scheduling stability penalty term of the i th idle robot for executing the to-be-assigned task j, a set of key resource points passed through by the future motion trajectory of the i th idle robot for executing the to-be-assigned task j , a delay conversion coefficient, a space-time aggregation density of the moment at which the i th idle robot reaches the key resource point when executing the to-be-assigned task j, a maximum value function.

[0023] According to the multi-layer shuttle robot dynamic task allocation method based on machine learning provided by the application, the key resource points comprise: a hoist inlet, a hoist outlet, a roadway intersection and a charging pile area.

[0024] The application can concentrate the calculation resources in the bottleneck area where congestion is most likely to occur, greatly reduce the calculation complexity, and improve the real-time performance of the algorithm on the premise of ensuring the effectiveness of the congestion prediction.

[0025] The method for dynamically assigning tasks to multi-layer shuttle robots based on machine learning provided by the application selects an idle robot with the minimum final evaluation value to execute a task to be assigned, and comprises the following steps of: a WCS sending a task execution instruction to a target robot with the minimum final evaluation value, the target robot updating a running state to be in operation, and the target robot performing a task along a future motion trajectory of the target robot as a path.

[0026] The application has the following advantages:

[0027] Based on the above technical solution, the method for dynamically assigning tasks to multi-layer shuttle robots based on machine learning provided by the application proposes a globally optimized dynamic task assignment method, which can effectively improve the task assignment accuracy of the multi-layer shuttle robots and improve the work efficiency. In the task assignment process, the method can accurately calculate the benchmark task time of the robot by comprehensively calculating the ideal execution time including the horizontal movement time and the dynamic waiting time predicted based on the number of elevators in the queue; on this basis, the application further introduces a scheduling stability penalty term based on the prediction of the future motion trajectory to evaluate potential traffic congestion, thereby constructing a more comprehensive task cost evaluation model. This method not only considers the execution efficiency of a single robot, but also takes into account the running harmony of the entire robot system, which can fundamentally avoid the elevator or critical path congestion problem caused by improper decision-making, realize the global optimization of the robot task assignment, and significantly improve the running efficiency and throughput of the entire warehouse system. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is a step flow chart of the method for dynamically assigning tasks to multi-layer shuttle robots based on machine learning provided by the embodiment of the application. DETAILED DESCRIPTION

[0029] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are part of the embodiments of the application, rather than all the embodiments of the application.

[0030] In order to improve the task assignment efficiency of cross-layer work in the multi-layer shuttle system, the embodiment of the application provides a method for dynamically assigning tasks to multi-layer shuttle robots based on machine learning, which is shown in Figure 1 which shows a step flow chart of the method for dynamically assigning tasks to multi-layer shuttle robots based on machine learning provided by the embodiment of the application, and the method comprises the following steps:

[0031] S1: obtaining task information to be assigned, state information of each elevator, and state information of all idle robots.

[0032] Specifically, the warehouse control system (WCS) can poll and integrate the state information of the three core entities of the shuttle robot, the task queue, and the elevator in real time, structure the discrete data, form a unified multi-dimensional dynamic information matrix, and provide real-time system data for subsequent data processing.

[0033] After receiving the new to-be-assigned task information, the WCS will broadcast the to-be-assigned task to the idle robots, and after selecting a suitable idle robot, the idle robot will immediately modify the current running state and execute the corresponding task.

[0034] For example, in the embodiment of the application, the to-be-assigned task information, the state information of each elevator, and the state information of all idle robots are obtained, and before this, the running data of the elevators and the robots are collected and preprocessed to obtain the state information of the elevators and the state information of the robots; wherein the state information of the elevators includes the number of robots queuing at the entrance of the elevator and the single service time in the historical service data of the elevator; the state information of the robots includes the current position, the running state, the running speed, and the actual power; the to-be-assigned task information at least includes the task point position, and the robot with an idle running state is regarded as an idle robot, and the running state includes idle, working, etc.

[0035] The to-be-assigned task information, the state information of each elevator, and the state information of all idle robots are obtained at each time. It should be understood that the real-time position, power, and queuing number of the dynamic data robot and the elevator are obtained through sensors or communication, and there may be instantaneous communication delay, sensor misreading or jitter. If not denoised, abnormal spikes may be included, resulting in low accuracy of subsequent data processing.

[0036] Based on this, the embodiment of the application can first preprocess the data before performing the subsequent data processing steps.

[0037] Since the robot may have communication interruption problems in remote sections or poor signal areas, missing values need to be identified and processed, so data cleaning and data calibration preprocessing of the robot data are required. In addition, since the collected various data come from heterogeneous data of different entities such as robots, elevators, and tasks, data structuring processing is required to integrate various data into structured data to provide a unified input format for subsequent algorithms. The specific implementation steps of preprocessing can be realized by existing technologies, and the embodiment of the application will not be repeated here.

[0038] According to the above steps, the to-be-assigned task information, the state information of each elevator and the state information of all idle robots can be obtained. However, in the multi-layer shuttling process, the queue number of the elevator obtained is a transient value, which cannot show the future trend, and therefore it is necessary to predict the interaction process between the robot and the elevator and plan and select the idle robot according to the prediction result to improve the efficiency and safety of the interaction, that is, the following steps are performed.

[0039] S2: For each combination of the to-be-assigned task and the idle robot, calculate the ideal execution time, which includes the horizontal movement time between the idle robot and the task point position and the dynamic waiting time required for cross-layer work predicted based on the number of queued robots and the historical service data of the elevator.

[0040] It should be noted that, in order to avoid the robot going to an elevator that may soon be congested, the present embodiment can predict the waiting time of the robot at the elevator, and by combining the horizontal movement time of the robot, determine the idle robot with less required ideal execution time as the target robot for executing the to-be-assigned task, thereby reducing the task time and improving the running efficiency.

[0041] It should be understood that in a warehouse environment, the robot usually moves along a fixed lane and path in a straight line, and therefore, when determining the horizontal movement time of the robot, the present embodiment can evaluate the distance between the idle robot and the task point position by Manhattan distance.

[0042] For example, in the present embodiment, the horizontal movement time of the idle robot is obtained in the following manner: the Manhattan distance between the idle robot from the current position to the task point position is obtained, and the ratio of the Manhattan distance to the running speed of the idle robot is taken as the horizontal movement time between the idle robot and the task point position.

[0043] The Manhattan distance can be determined by a shortest path planning algorithm, which can be set according to actual needs.

[0044] For example, in the present embodiment, the dynamic waiting time required for cross-layer work predicted based on the number of queued robots and the historical service data of the elevator includes: the product of the predicted waiting time of the last period and the actual waiting time of the current period is weighted and summed to obtain the dynamic waiting time required for cross-layer work, wherein the actual waiting time is the product of the current number of queued robots of the elevator and the average single service time in the historical service data.

[0045] For ease of understanding, the present embodiment provides a calculation method of the dynamic waiting time required for cross-layer work predicted based on the number of queued robots and the historical service data of the elevator, which can be seen from the following relationship:

[0046] ;

[0047] For the current period The Dynamic waiting time required for cross-level operations of a single hoist. As a smoothing factor, For the first period of the previous cycle Dynamic waiting time required for cross-level operations of a single hoist. This represents the current number of queuing robots for the current cycle lift. For the current cycle number The average service time per session is based on the historical service data of each hoist.

[0048] In the above formula, Indicating historical trends The weight used to determine the current period can be set to 0.5. The weight of the previous period, For the current cycle number Total service time required for each hoist.

[0049] Based on the above steps, the dynamic waiting time required for cross-level operations of each hoist can be obtained. By obtaining the hoists on the path where the idle robot executes the assigned task, the dynamic waiting time required for the corresponding idle robot to execute the assigned task for cross-level operations can be obtained. Combined with the horizontal movement time between the idle robot and the task point location, and the dynamic waiting time required for cross-level operations predicted based on the number of queuing robots and the historical service data of the hoists, the ideal execution time can be determined.

[0050] It should be further noted that if the idle robot has low battery power while performing a task, it may run out of power midway, affecting the efficiency of task execution and possibly even causing a blockage.

[0051] Based on this, in determining the ideal execution time for each idle robot to perform the assigned task, the embodiments of the present invention can also weight the time cost by the battery level of the idle robot. By increasing the time cost of the idle robot with low battery level, the likelihood of it being selected as the robot to perform the assigned task is reduced, the battery life of the low-battery robot is extended, and charging time is gained for it.

[0052] To reduce the possibility of congestion, a preset battery threshold can be set to force users with low battery levels to exit the current screening process.

[0053] For example, in an embodiment of the present invention, if the actual battery level of the idle robot is less than the battery level threshold, the idle robot is forced to perform a charging task.

[0054] The electric quantity threshold can be set as 20% of the full electric quantity, and can be set according to actual needs.

[0055] For example, in the embodiment of the application, the ideal execution time is calculated, and the following relationship can be referred to:

[0056]

[0057] is the ideal execution time of the i th idle robot for executing the to-be-assigned task j, is the horizontal movement time between the current position of the i th idle robot and the task point position of the to-be-assigned task j, is the dynamic waiting time required by the i th idle robot for cross-layer operation of the to-be-assigned task j, is the full electric quantity of the idle robot, is the actual electric quantity of the i th idle robot.

[0058] In this calculation method, is the direct time cost, and the greater the value, the longer the time required by the idle robot for executing the to-be-assigned task, and thus the lower the possibility of being selected as the target robot.

[0059] is the electric quantity adjustment factor, and the greater the difference between the full electric quantity and the actual electric quantity of the idle robot, the smaller the actual electric quantity of the idle robot, and thus the greater the electric quantity adjustment factor is required to increase the time cost and reduce the possibility of being selected as the target robot.

[0060] Therefore, the greater the direct time cost of the idle robot for executing the to-be-assigned task, and the smaller the electric quantity, the greater the possibility that the idle robot is not suitable as the target robot, and the ideal execution time of each idle robot for executing the to-be-assigned task can be obtained according to the above steps. When the target robot for executing the to-be-assigned task is selected from the idle robots, the greater the ideal execution time of the idle robot, the longer the time required, the higher the corresponding bidding cost, and the lower the possibility of being selected as the target robot.

[0061] S3: predicting future motion trajectories of all robots, and calculating a scheduling stability penalty term of the idle robot.

[0062] The scheduling stability penalty term is positively correlated with the spatiotemporal aggregation density of each key resource point on the future motion trajectory of the idle robot.

[0063] ​It should be noted that the ideal execution time obtained based on the above steps can evaluate the time-consuming degree of the idle robot, and the task allocation of the robot can be realized according to the time-consuming degree. However, the robot is not an ideal particle, and it will have a potential traffic influence on the surrounding area during movement. If the best target robot is directly determined according to the time-consuming degree, a scheduling scheme in which all robots rush to the same hoist or the same key intersection at the same time may occur, so that the actual completion time of the task is much greater than the ideal execution time, and the task execution efficiency is reduced.

[0064] Based on this, the embodiment of the present application can construct a space-time occupation potential field model for representing the influence degree of the robot on the space point at each time in the future motion trajectory, to describe the space-time influence of the robot at the future time, so as to convert the future motion trajectory of each robot into a continuous and time-varying traffic pressure field. By superimposing the potential field strength of all robots at the same time and the same position, the space-time occupation density of the space-time point can be obtained, and the priority of each idle robot can be evaluated based on this, which can effectively reduce the possibility of congestion or resource contention at the key resource point.

[0065] For example, in the embodiment of the present application, the ideal execution time can be modified by calculating the scheduling stability penalty term of the idle robot to obtain the final evaluation value of each idle robot.

[0066] It should be understood that by predicting the future motion trajectory of all robots, the robot passing through the key resource point at each future time can be obtained, and by superimposing the space-time occupation potential field of the robot passing through the key resource point at each future time, the density of the key resource point at the future time can be accurately obtained.

[0067] Specifically, at the key resource point passed through by the future motion trajectory, the space-time occupation potential fields generated by all robots on the key resource point are superimposed to obtain the space-time aggregation density of the key resource point; the scheduling stability penalty term of the idle robot is calculated, and the scheduling stability penalty term is positively correlated with the space-time aggregation density of each key resource point on the future motion trajectory of the idle robot.

[0068] Among them, the future motion trajectory of all idle robots when executing the to-be-assigned task can be predicted according to the path planning algorithm, which can be realized by the prior art, and the embodiment of the present application will not be repeated here.

[0069] For example, in the embodiment of the present application, the key resource point in the space point can be set as the hoist inlet, the hoist outlet, the roadway intersection and the charging pile area, etc., which can be set according to actual needs.

[0070] It should be understood that the space point contains the key resource point, and the key resource point is a kind of space point.

[0071] For example, in an embodiment of the present invention, the method for obtaining the spatiotemporal occupancy potential field generated by the robot includes: obtaining the predicted position of the robot at each moment in its future motion trajectory; constructing a spatiotemporal occupancy potential field model to characterize the degree of influence of the robot on the spatial point at each moment in its future motion trajectory, thereby obtaining the spatiotemporal occupancy potential field generated by the robot on the spatial point; wherein, the spatiotemporal occupancy potential field model is a Gaussian potential field model, and the spatiotemporal occupancy potential field has an inverse exponential relationship with the square of the distance from the spatial point to each predicted position of the robot.

[0072] For ease of understanding, embodiments of the present invention provide a method for constructing a spatiotemporal occupancy potential field model, as detailed in the following relationships:

[0073] ;

[0074] For the first A robot in its future trajectory At any time, the spatial point The resulting spatiotemporal occupancy potential field, For spatial points Location, For the first A robot in its future trajectory Predicted location at time Represents the square of the distance. The radius of influence of the potential field. It is an exponential function with base e.

[0075] The potential field influence radius is used to control the potential field influence range of the robot. It can be set by the robot's radius, safety distance, and lane width. The specific value can be determined according to actual needs. This embodiment of the invention does not impose too many restrictions here.

[0076] In the above formula, if the robot's future trajectory... The closer the predicted position at a given moment is to the position of a spatial point, the greater the influence of the robot on the potential field of that spatial point.

[0077] Substituting the key resource points into the aforementioned spatiotemporal occupancy potential field model, we can obtain the spatiotemporal occupancy potential field generated by each robot at each moment in its future trajectory. By superimposing the spatiotemporal occupancy potential fields generated by all robots at key resource points, we can obtain the spatiotemporal clustering density of key resource points at each future moment. The higher the spatiotemporal clustering density, the greater the probability of congestion. Based on the spatiotemporal clustering density, we can construct a scheduling stability penalty term for each idle robot, which can accurately determine the delay time required for congestion.

[0078] To reduce the amount of data processing, in this embodiment of the invention, robots on the future motion trajectory of idle robots can also be selected for spatiotemporal occupancy potential field calculation.

[0079] Specifically, in this embodiment of the invention, the future running time of the idle robot is determined based on the future trajectory length of the idle robot executing the assigned task and the robot's running speed; within the future running time, the spatiotemporal occupancy potential field of each robot passing through the key resource point on the future trajectory of the idle robot is obtained, and the spatiotemporal occupancy potential fields generated by all robots on the key resource point are superimposed to obtain the spatiotemporal clustering density of the key resource point.

[0080] For example, in an embodiment of the present invention, the scheduling stability penalty term for the idle robot is calculated, as shown in the following formula:

[0081] ;

[0082] The scheduling stability penalty term is applied to the i-th idle robot for executing the task j to be assigned. For the i-th idle robot to perform the assigned task j, the key resource points it passes through on its future trajectory. The set, For the delay conversion factor, The time when the i-th idle robot arrives at the critical resource point while performing the assigned task j. The spatiotemporal aggregation density, This is the function for finding the maximum value.

[0083] The delay conversion coefficient is in seconds and is used to convert the dimensionless congestion index into actual delay time. As an example, in this embodiment of the invention, the delay conversion coefficient can be set to 0.1 seconds, which can be set according to actual needs.

[0084] In the above formula, when the idle robot arrives at the critical resource point, if the spatiotemporal clustering density of that critical resource point is greater than 1, that is... A value greater than 0 indicates that congestion may occur near the key resource point, and therefore it is necessary to penalize the possibility of the idle robot being selected as the target robot.

[0085] Conversely, when the spatiotemporal clustering density of the key resource point is less than or equal to 1, i.e. A value less than or equal to 0 indicates that the idle robot will not cause congestion near the critical resource point when it arrives at the critical resource point. Therefore, no penalty is required, and the value can be set directly to 0 by taking the maximum value.

[0086] According to the above steps, the scheduling stability penalty item of each idle robot performing the to-be-assigned task can be accurately obtained, and by combining the delay time of the scheduling stability penalty item with the ideal execution time, the possibility of each idle robot being selected as the target robot can be accurately obtained.

[0087] S4: adding the ideal execution time of the idle robot and the scheduling stability penalty to obtain a final evaluation value; and selecting the idle robot with the minimum final evaluation value to perform the to-be-assigned task.

[0088] It should be noted that the ideal execution time representing individual efficiency and the scheduling stability penalty representing system influence can be obtained according to the above steps, and the embodiment of the present application combines the two to construct a final evaluation value, which can comprehensively reflect the individual efficiency and global harmony of the idle robot, so as to accurately select the best idle robot.

[0089] For example, in the embodiment of the present application, selecting the idle robot with the minimum final evaluation value to perform the to-be-assigned task comprises: the WCS sends a task execution instruction to the idle robot with the minimum final evaluation value as the target robot, the target robot updates the running state to working, and the future motion trajectory of the target robot is taken as a path for performing the to-be-assigned task.

[0090] For example, in the embodiment of the present application, after the task assignment is implemented, the future motion trajectory of the idle robot can also be taken as a standard path into the system to provide a basis for subsequent path selection of the robot.

[0091] For example, in the embodiment of the present application, after the target robot completes the to-be-assigned task, the actual waiting time of the target robot in performing the to-be-assigned task can also be taken as an index into the calculation of the final evaluation value, which can be set according to actual needs, and the embodiment of the present application will not be repeated here.

[0092] It can be seen that in the embodiment of the application, when the task allocation of the robot is implemented, the to-be-allocated task information, the state information of each elevator and the state information of all idle robots can be acquired, the state information of the elevator includes the number of queued robots at the entrance of the elevator; for each combination of the to-be-allocated task and the idle robot, the ideal execution time is calculated, the ideal execution time includes the horizontal movement time between the idle robot and the task point position and the dynamic waiting time required for cross-layer work predicted based on the number of queued robots and the historical service data of the elevator; the future motion trajectory of the robot is predicted, at the key resource points passed through by the future motion trajectory, the space-time occupation potential field generated by all robots on the key resource points is superimposed to obtain the space-time aggregation density of the key resource points; the scheduling stability penalty term of the idle robot is calculated, the scheduling stability penalty term is positively correlated with the space-time aggregation density of each key resource point on the future motion trajectory of the idle robot; the ideal execution time of the idle robot and the scheduling stability penalty are time-added to obtain the final evaluation value; the idle robot with the minimum final evaluation value is selected to execute the to-be-allocated task, which effectively improves the accuracy of the robot task allocation and improves the efficiency of the robot in executing the task.

[0093] The above merely describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. within the principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for dynamic task allocation of multi-layered shuttle robots based on machine learning, characterized in that, The method comprises the following steps: Collecting and preprocessing the operation data of the elevators and the robots to obtain state information of the elevators and state information of the robots; the state information of the elevators comprises single-service time in historical service data of the elevators and the number of queued robots at the entrances of the elevators; the state information of the robots comprises current positions, running states, running speeds and actual power; Obtaining task information to be allocated, state information of each elevator and state information of all idle robots; the task information to be allocated at least comprises the positions of task points, and the robots in an idle state are regarded as idle robots; For each combination of the task to be allocated and the idle robots, calculating ideal execution time, which comprises horizontal movement time between the idle robots and the positions of the task points and dynamic waiting time required for cross-layer work predicted based on the number of queued robots and historical service data of the elevators, and meeting the following condition: ； , respectively are the ideal execution time of the i-th idle robot to perform the task j to be assigned, the dynamic waiting time required for cross-layer jobs, is the horizontal movement time between the current position of the i-th idle robot and the task point position to perform the task j to be assigned, is the full battery capacity of the idle robot, is the actual battery capacity of the i-th idle robot; Predicting future motion trajectories of the robots, superimposing time-space occupation potential fields generated by all the robots on key resource points in the future motion trajectories to obtain time-space aggregation densities of the key resource points; Calculating a scheduling stability penalty term of the idle robots, which is positively correlated with the time-space aggregation densities of each key resource point in the future motion trajectory of the idle robots: ； , respectively are the scheduling stability penalty term for the ith idle robot to execute the task j to be assigned, the set of critical resource points passed by the future motion trajectory of the ith idle robot to execute the task j to be assigned, is the delay conversion coefficient, is the spatiotemporal concentration of the time instants at which the ith idle robot arrives at the critical resource points when executing the task j to be assigned, is the max function;​​ Adding the ideal execution time of the idle robots and the scheduling stability penalty in time to obtain a final evaluation value; and selecting an idle robot with the minimum final evaluation value to execute the task to be allocated.

2. The machine learning based dynamic task allocation method for multi-tiered shuttle robots according to claim 1, wherein, The horizontal movement time of the idle robots is obtained in the following manner: Obtaining the Manhattan distance between the current position of the idle robot and the position of the task point, and taking the ratio of the Manhattan distance and the running speed of the idle robot as the horizontal movement time between the idle robot and the position of the task point.

3. The machine learning based multi-tiered shuttle robot dynamic task assignment method of claim 1, wherein, The dynamic waiting time required for cross-layer work predicted based on the number of queued robots and historical service data of the elevators comprises the following steps: Weighted sum of the product of the predicted waiting time in the last period and the actual waiting time in the current period to obtain the dynamic waiting time required for cross-layer work, wherein the actual waiting time is the product of the current number of queued robots of the elevator and the average single-service time in the historical service data.

4. The machine learning based multi-tiered shuttle robot dynamic task assignment method of claim 1, wherein, The future motion trajectories of all the idle robots are predicted according to a path planning algorithm.

5. The machine learning based multi-tiered shuttle robot dynamic task assignment method of claim 1, wherein, The time-space occupation potential field generated by the robot is obtained in the following manner: Obtaining the predicted positions of the robot at each time in the future motion trajectory; constructing a time-space occupation potential field model for representing the influence degree of the robot on a space point at each time in the future motion trajectory to obtain the time-space occupation potential field generated by the robot on the space point; wherein the time-space occupation potential field model is a Gaussian potential field model, and the time-space occupation potential field is inversely exponentially related to the square of the distance from the space point to each predicted position of the robot.

6. The machine learning based multi-tiered shuttle robot dynamic task assignment method of claim 1, wherein, The key resource points comprise the entrances of the elevators, the exits of the elevators, the intersections of the passageways and the charging pile areas.

7. The machine learning based multi-tiered shuttle robot dynamic task assignment method of claim 1, wherein, The idle robot with the minimum final evaluation value is selected to execute the task to be allocated in the following manner: The WCS sends a task execution instruction to a target robot with the smallest final evaluation value among the idle robots, and the target robot updates a running state to working and performs a task along a path of a future motion trajectory of the target robot.

Images