CN112987721B - Multi-AGV scheduling device and fusion method of global planning and local planning thereof - Google Patents

Abstract

The invention discloses a multi-AGV scheduling device and a fusion method of global planning and local planning thereof. The application layer provides an interface for developers to use an AGVTroop framework; the AGVTroop framework completes task allocation and path planning according to the issued tasks, the map information and the AGV information; the ROS layer is used for constructing a communication frame, controlling bottom layer equipment and combining a software package provided by the ROS and laser information of the AGV to achieve positioning of the AGV. The invention adopts a method of fusing global planning and local planning. Firstly, a route with global optimality is planned according to global indexes based on an improved A-x algorithm, and a local planner dynamically adjusts the global route according to local information and designated strategies and indexes.

Description

A multi-AGV scheduling device and its global planning and local planning fusion method

technical field

The invention belongs to the field of AGVs, and in particular relates to a multi-AGV dispatching device and a fusion method of global planning and local planning thereof.

Background technique

The global planning of multiple AGVs will refer to the driving paths and states of other AGVs when planning the path for the current AGV, and consider global indicators such as path balance and total task completion time. Common global planning methods include single-item graph method and traffic method.

The one-way map method simplifies the AGV driving model by specifying a single driving direction of the route. The possible driving path in the map is single, and conflicts can only occur at intersections, which is convenient for the design of conflict avoidance strategies. However, since the allowable directions of all paths are stipulated, detours are prone to occur, and the path cost of task execution is relatively high.

Traffic law is mainly used for resource allocation at intersections. Collisions are caused by multiple objects occupying a route at the same time. Therefore, each collision can be avoided by taking appropriate measures for all intersections in the relevant environment. By enumerating overlapping routes: routes in opposite directions, routes in the same direction, and intersections, construct traffic rules that accommodate all these overlapping routes to avoid collisions. Common processing methods include: intersection waiting, path reselection, etc. Although the traffic law can effectively resolve conflicts and reduce the incidence of accidents, it is more about strategic handling when conflicts are about to occur. When the number of AGVs is large and the traffic congestion is serious, it is easy to cause system paralysis.

The global planning algorithm considers the global optimality, which can reduce the processing time of tasks to a certain extent, and avoid problems such as system deadlock and collision. However, the time complexity of the global planning algorithm increases exponentially with the increase of AGV, the response time is long, and the real-time performance is low, which makes it difficult to handle the dynamic scene of AGV.

Dijkstra's algorithm spreads out the adjacent nodes with the starting point as the center, uses the cumulative cost of the search path as a heuristic, and preferentially expands the node with the smallest cumulative cost of the path. The A* algorithm is an algorithm obtained by combining the branch and bound method and the expansion list tool with admissible heuristics. While considering the cumulative path cost, it estimates the lower bound of the remaining path cost of the expansion node, avoids the search from going in the opposite direction, and accesses Expand the list to avoid repeated expansion.

The local planning of AGV only considers the path planning of the current AGV. The time complexity will not increase with the number of AGVs, and it will remain within the calculable range. The response time is short, the real-time performance is high, and the dynamic scene of AGV can be well handled. However, since AGV only considers local optimality in path planning, when the number of AGVs increases, problems such as deadlocks and collisions will frequently occur.

Contents of the invention

The present invention provides a multi-AGV dispatching device and a global planning and local planning fusion method thereof, which are used to solve the above problems and can flexibly cope with different scenarios.

The present invention is realized through the following technical solutions:

A multi-AGV scheduling device, the device includes an application layer, an AGVTroop framework, a ROS layer and a plurality of AGVs, and the application layer provides developers with an interface to use the AGVTroop framework;

The AGVTroop framework completes task assignment and path planning according to released tasks, map information, and AGV information;

The ROS layer builds a communication framework, controls the underlying equipment, and realizes the positioning of the AGV in combination with the software package provided by ROS and the laser information of the AGV.

Further, the communication framework realizes communication between AGVCluster and AGV;

The underlying device is an actual AGV.

A global planning and local planning fusion method for multiple AGV scheduling devices, the fusion method comprising the following specific steps:

Step 1: Read the information of the global planner based on the improved A* algorithm, that is, the time window information;

Step 2: The local planner performs local path planning with reference to the global route according to the local information, and outputs the speed command to the time window information;

Step 3: Determine whether the local planning time exceeds the threshold, if it exceeds the threshold, return to the global planner in step 1, and if it does not exceed the threshold, proceed to step 4;

Step 4: Determine whether the trajectory with the highest score is less than 0, if it is less than 0, return to the global planner in step 1, and if it is greater than 0, output the speed command.

Further, the reference indicators in the global planner include all task completion times and path balance;

The reference indicators in the local planner include deviation from the global route, direction cost and obstacle cost.

Further, the global planning method in the global planner in step 1 specifically includes the following steps:

Step Q1: Receive the task to be executed, if the task to be executed is empty, enter the end, if the task to be executed is not empty, enter step Q2;

Step Q2: Construct a direction time window and an occupancy time window for each node of each grid map;

Step Q3: Use the direction time window and occupied time window constructed in step Q2 to call the improved A* algorithm for global planning;

Step Q4: Utilize the global planning in step Q3 to update the direction time window and occupied time window of nodes on the path;

Step Q5: According to step Q4, it is judged whether there is a node direction conflict, if yes, go to step Q6 and Q7, if not, go to step Q8;

Step Q6: Calculate the time cost of waiting;

Step Q7: re-plan the path;

Step Q8: According to step Q4, it is judged whether there is a node occupation time window conflict, if yes, go to step Q6, if not, go to step Q9;

Step Q9: Send the path to the corresponding AGV;

Step Q10: According to step Q6 and step Q7, judge whether the waiting time cost is greater than the re-planning time, if it is greater, proceed to step Q11, and if it is less, proceed to step Q12;

Step Q11: Update the global path, and send the path to the corresponding AGV;

Step Q12: Wait in place and send the path to the corresponding AGV.

Further, the local planning method specifically includes the following steps:

Step J1: Obtain the global planning route;

Step J2: Read the lidar data based on the global planning route in step J1 to obtain local information;

Step J3: Sampling multiple groups of velocities in the velocity space (v, w) based on the local information in step J2;

Step J4: Simulate its trajectory within 0.01s-0.10s based on multiple sets of sampling speeds in step J3;

Step J5: scoring the trajectory of step J4;

Step J6: Judging whether the score of step J5 is greater than the threshold, that is, whether there is a feasible path, if the score is greater than the threshold, enter step J7, and if the score is less than the threshold, enter step J8;

Step J7: Execute according to the trajectory with the highest score;

Step J8: Correct the direction time window, occupied time window and usage frequency of nodes on the trajectory;

Step J9: Dynamically adjust the global planning route based on the corrected information in step J8.

The beneficial effects of the present invention are:

In order to ensure the global optimality of AGV planning and improve the real-time performance of the system and the ability to process dynamic scenes, the present invention adopts a fusion method of global planning and local planning. First, based on the improved A* algorithm, a route with global optimality is planned according to the global index, and the local planner will dynamically adjust the global route according to the specified strategy and index according to the local information.

Self-tuning framework, in order to make the framework flexible to respond to different scenarios, the system provides an extensible platform to support different optimization rules as plug-ins. The system supports multiple optimization models, and developers can customize the performance indicators of new task scheduling or path gauges, and the newly added indicators are used as plug-ins. The AGVTroop framework needs to achieve self-optimization, and adjust the allocation scheduling strategy and AGV path planning according to the newly added indicators. And select the best weighted combination for these performance indicators by replaying historical data.

The present invention provides a fine-grained interface, supports action-level scheduling, enables developers to program task-oriented, and only needs to customize tasks and task groups and publish them, without caring about how tasks are executed, simplifying AGV application development.

Description of drawings

Fig. 1 is a schematic diagram of the overall architecture of the present invention.

Fig. 2 is a schematic diagram of the application layer of the present invention.

Fig. 3 is a schematic diagram of the AGVTroop layer of the present invention.

Fig. 4 is a schematic diagram of the ROS layer of the present invention.

Fig. 5 is a flow chart of the integration of global planning and local planning in the present invention.

Fig. 6 is a flowchart of global path planning in the present invention.

Fig. 7 is a flowchart of local path planning in the present invention.

detailed description

The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

A multi-AGV scheduling device, the device includes an application layer, an AGVTroop framework, a ROS layer, and a plurality of AGVs, the application layer provides developers with an interface to use the AGVTroop framework; developers only need to issue tasks according to application scenarios, and do not need to care How to perform tasks and which AGVs will be assigned to them;

The AGVTroop framework completes task assignment and path planning according to released tasks, map information, and AGV information;

The ROS layer builds a communication framework, controls the underlying equipment, and realizes the positioning of the AGV in combination with the software package provided by ROS and the laser information of the AGV.

Further, the communication framework realizes communication between AGVCluster and AGV;

The underlying device is the actual AGV; compared to the AGVCluster, the AGVCluster is oriented towards virtual AGV scheduling, and the ROS completes the binding of the virtual AGV and the actual AGV and controls the underlying device.

A global planning and local planning fusion method for multiple AGV scheduling devices, the fusion method comprising the following specific steps:

Step 1: Read the information of the global planner based on the improved A* algorithm, that is, the time window information;

Step 2: The local planner performs local path planning with reference to the global route according to the local information, and outputs the speed command to the time window information;

Step 3: Determine whether the local planning time exceeds the threshold, if it exceeds the threshold, return to the global planner in step 1, and if it does not exceed the threshold, proceed to step 4;

Step 4: Determine whether the trajectory with the highest score is less than 0, if it is less than 0, return to the global planner in step 1, and if it is greater than 0, output the speed command.

Further, the reference indicators in the global planner include all task completion times and path balance;

The reference indicators in the local planner include deviation from the global route, direction cost and obstacle cost.

Further, the global planning method in the global planner in step 1 specifically includes the following steps:

Step Q1: Receive the task to be executed, if the task to be executed is empty, enter the end, if the task to be executed is not empty, enter step Q2;

Step Q2: Construct a direction time window and an occupancy time window for each node of each grid map;

Step Q3: Use the direction time window and occupied time window constructed in step Q2 to call the improved A* algorithm for global planning;

Step Q4: Utilize the global planning in step Q3 to update the direction time window and occupied time window of nodes on the path;

Step Q5: According to step Q4, it is judged whether there is a node direction conflict, if yes, go to step Q6 and Q7, if not, go to step Q8;

Step Q6: Calculate the time cost of waiting;

Step Q7: re-plan the path;

Step Q8: According to step Q4, it is judged whether there is a node occupation time window conflict, if yes, go to step Q6, if not, go to step Q9;

Step Q9: Send the path to the corresponding AGV;

Step Q10: According to step Q6 and step Q7, judge whether the waiting time cost is greater than the re-planning time, if it is greater, proceed to step Q11, and if it is less, proceed to step Q12;

Step Q11: Update the global path, and send the path to the corresponding AGV;

Step Q12: Wait in place and send the path to the corresponding AGV.

Further, the local planning method specifically includes the following steps:

Step J1: Obtain the global planning route;

Step J2: Read the lidar data based on the global planning route in step J1 to obtain local information;

Step J3: Sampling multiple groups of velocities in the velocity space (v, w) based on the local information in step J2;

Step J4: Simulate its trajectory in a short period of time (the default value is 0.05s) based on the multiple sets of sampling speeds in step J3;

Step J5: scoring the trajectory of step J4;

Step J6: Judging whether the score of step J5 is greater than the threshold, that is, whether there is a feasible path, if the score is greater than the threshold, enter step J7, and if the score is less than the threshold, enter step J8;

Step J7: Execute according to the trajectory with the highest score;

Step J8: Correct the direction time window, occupied time window and usage frequency of nodes on the trajectory;

Step J9: Dynamically adjust the global planning route based on the corrected information in step J8.

Further, the improved A* algorithm specifically includes adding global information when calculating node n in the search path, and the calculation formula f(n) is:

f(n)=g(n)+h(n)+α*Math.min(time_threshold, cur_time-last_time)+β*frequency[n]+δ*steering_correction

Among them, g(n): the cost from the starting node to node n; h(n): the estimated cost of the best path from node n to the target node; time_threshold: the maximum time threshold; cur_time: the current time of using node n; last_time: last use time of point n; frequency[n]: usage frequency of node n; steering_correction: steering correction; α, β, δ are time weight, frequency weight and steering weight respectively;

The path planning of the present invention adopts the fusion mode of global planning and local planning. Based on the improved A* algorithm, a route with global optimality will be planned according to a series of global indicators (time of completion of all tasks, path balance, etc.), and local planning The controller will dynamically adjust the global route according to the local information. (Reference indicators include deviation from the global route, direction cost, obstacle cost, etc. Adjustment strategies include route adjustment, static waiting, re-planning, etc.).

The present invention provides a fine-grained interface, supports action-level scheduling, and enables developers to program task-oriented. They only need to customize tasks and task groups and publish them, without caring about how tasks are executed, simplifying AGV application development. ROS provides various types of hardware abstraction and drivers, so that the system can understand the differences of the underlying AGV, and make the system oriented to virtual AGV programming.

In order to make the framework flexible to respond to different scenarios, the system provides an extensible platform to support different optimization rules as plug-ins. The system supports multiple optimization models, and developers can customize the performance indicators of new task scheduling or path gauges, and the newly added indicators are used as plug-ins. The AGVTroop framework needs to achieve self-optimization, and adjust the allocation scheduling strategy and AGV path planning according to the newly added indicators. And choose the best weighted combination for these performance indicators.

The global path planning adopts the A* algorithm based on the time window to improve the multi-AGV path planning, and constructs the direction time window and the occupation time window for each node of the grid map. The global path is calculated based on the A* algorithm, and the occupation time window and direction time window of each node on the path are calculated according to the hardware information of the AGV. The A* algorithm needs to plan the path according to the time window information, considering three factors (1) Accessibility, according to the current time window information, there must be no node on the constructed path that causes conflicts in direction or occupied time. (2) Collision avoidance, consider the frequency of nodes being occupied within the arrival time when planning the path, and try to avoid path imbalance (that is, there are multiple AGVs accessing a certain channel, while other channels are idle) (3) Steering correction, try to plan the path Avoid changing the driving direction. Frequent steering will not only seriously increase the energy consumption of the AGV, aggravate the wear and tear of the hardware, but also easily cause AGV conflicts. Finally, the global route is sent to the local planner for adjustment. (4) Time cost, the time required to complete all tasks.

Local path planning is mainly used to dynamically adjust the route of the global plan. Although global planning considers the global optimum, it cannot handle dynamic scenes well (such as obstacles that are not on the newly added map, hardware failures, and dynamic objects such as people). Local path planning fine-tunes the global route based on the local information scanned by the lidar. Partial path planning adopts DWA dynamic window algorithm. Sampling multiple groups of speeds in the speed space (v, w) (the speed can be controlled within a certain range according to the limitations of the car itself and the environment when sampling the speeds), and simulate the trajectory of the car at these speeds for a certain period of time, and get multiple After grouping trajectories, evaluate these trajectories, and select the speed group corresponding to the optimal trajectory to drive the robot to move. The scoring function is a weighted sum of the following costs.

(1) Shock cost, whether it is shocking or not, that is, switching the walking direction multiple times in a short period of time.

(2) Obstacle cost, the distance between the trajectory and the obstacle.

(3) Path cost, the deviation value from the global path.

(4) Target cost, whether the direction is moving towards the target.

(5) Reverse cost, which punishes the backward path.

Claims (3)

1. A fusion method of global planning and local planning of a multi-AGV scheduling device is characterized by comprising an application layer, an AGVTroop frame, an ROS layer and a plurality of AGVs, wherein the application layer provides an interface for developers to use the AGVTroop frame;

the AGVTroop framework completes task allocation and path planning according to the issued tasks, the map information and the AGV information;

the ROS layer builds a communication frame, controls bottom layer equipment and combines a software package provided by the ROS and laser information of the AGV to realize positioning of the AGV;

the communication framework realizes the communication between the AGVCluster and the AGV; the bottom layer equipment is an actual AGV;

the fusion method comprises the following specific steps:

step 1: reading the information of the global planner based on the improved A-algorithm, namely time window information;

step 2: the local planner carries out local path planning by referring to the global route according to the local information and outputs a speed instruction to the time window information;

and step 3: judging whether the local planning time exceeds a threshold value, if so, returning to the global planner in the step 1, and if not, performing the step 4;

and 4, step 4: judging whether the track with the highest score is smaller than 0, if so, returning to the global planner in the step 1, and if so, outputting a speed instruction;

the global planning method in the global planner in step 1 specifically includes the following steps:

step Q1: receiving a task to be executed, if the task to be executed is empty, ending the process, and if the task to be executed is not empty, entering a step Q2;

step Q2: constructing a direction time window and an occupied time window for each node of each grid map;

and step Q3: calling an improved A algorithm by using the direction time window and the occupied time window constructed in the step Q2 to perform global planning;

step Q4: updating the direction time window and the occupied time window of the nodes on the path by using the global planning of the step Q3;

and step Q5: judging whether the node direction conflict exists according to the step Q4, if so, performing the steps Q6 and Q7, and if not, performing the step Q8;

and step Q6: calculating the waiting time cost;

step Q7: re-planning the path;

step Q8: judging whether the occupied time window conflict of the nodes exists according to the step Q4, if so, performing a step Q6, and if not, performing a step Q9;

step Q9: issuing the path to a corresponding AGV;

step Q10: judging whether the waiting time cost is larger than the re-planning time or not according to the step Q6 and the step Q7, if so, performing a step Q11, and if not, performing a step Q12;

step Q11: updating the global path and issuing the path to the corresponding AGV;

step Q12: waiting in place and issuing the path to the corresponding AGV;

the improved a-x algorithm is specifically that global information is added when a path calculation node n is searched, and a calculation formula f (n) is as follows:

f(n)＝g(n)+h(n)+α*Math.min(time_threshold,cur_time-last_time)+β*frequency[n]+δ*steering_correction

wherein g (n) is the cost from the starting node to node n; h (n): an estimated cost of the best path from node n to the target node; time _ threshold maximum time threshold; cur _ time: the time node n is currently in use; last _ time: point n last use time; frequency [ n ]: the frequency of use of the node n; curing _ correction: correcting the steering; alpha, beta and delta are respectively time weight, frequency weight and steering weight;

2. the method of claim 1, wherein the reference indicators in the global planner include the completion time of all tasks and the path balance;

the reference indicators within the local planner include a deviation from a global route value, a directional cost, and an obstacle cost.

3. The method for fusing global planning and local planning for multiple AGV schedulers according to claim 1, wherein said local planning method specifically comprises the following steps:

step J1: acquiring a global planning route;

step J2: reading laser radar data to obtain local information based on the global planning route of the step J1;

step J3: sampling a plurality of sets of velocities in a velocity space (v, w) based on the local information of step J2;

step J4: simulating the track of the multi-group sampling speed in the step J3 within 0.01s-0.10 s;

step J5: scoring the trajectory of step J4;

step J6: judging whether the score in the step J5 is larger than a threshold value, namely whether a feasible path exists, if so, entering a step J7, and if not, entering a step J8;

step J7: executing according to the track with the highest score;

step J8: correcting a direction time window, an occupied time window and a use frequency of a node on a track;

step J9: and D, dynamically adjusting the globally planned route based on the information corrected in the step J8.

Images