Multi-robot full-coverage search method based on improved bionic neural network

Abstract

The invention discloses a multi-robot full-coverage search method based on an improved bionic neural network, and the method can help a robot to achieve the effective obstacle avoidance through an improved bionic neural network model, and can also better attract the robot to search for an unknown search region. And meanwhile, the coverage search strategy adopts a mode of combining model prediction control and an auxiliary search mechanism, so that the robot can be timely helped to break away from a local deadlock state on the basis of effectively reducing the average turning frequency and the average turning angle of the robot, and the coverage search efficiency is greatly improved. The average single-step decision time for realizing full-coverage search by multiple robots is millisecond level, and the requirements in practical application can be met.

Description

Technical Field

[0001] The invention relates to the technical field of robot area coverage search, and is particularly suitable for a multi-robot full coverage search method based on an improved bionic neural network. Background Art

[0002] With the rapid development of robotics technology, people's demand for robot collaboration is also increasing. Robots are no longer just tools to complete a single task, but through division of labor and collaboration, they play a role in complex scenarios such as disaster relief, agricultural detection, warehouse management, environmental perception and target search, improving the efficiency and accuracy of task completion. The challenging problem of coverage search in unknown environments is how to maximize the exploration and coverage of environmental areas through the collaborative work of multiple robots and the sensors carried by the robots, so as to obtain comprehensive and accurate environmental information.

[0003] Although a variety of solutions have been proposed, there are still certain limitations. First, the uncertainty of the environment makes it difficult for robots to ensure full coverage when planning paths. Secondly, the collaboration and information sharing mechanisms between robots also need to be further optimized to improve the overall system efficiency. In addition, many existing methods still have deficiencies in real-time and adaptability, and cannot effectively cope with various situations in complex environments. Therefore, in unknown and complex terrain environments, robots need to have real-time obstacle avoidance capabilities to ensure that they can effectively plan paths. At the same time, they also need to have the ability to achieve full coverage of unknown areas with minimal energy consumption, optimize resource utilization, and improve overall work efficiency. Summary of the invention

[0004] The present invention aims to provide a multi-robot full coverage search method based on an improved bionic neural network, aiming to solve the problem of how to conduct full coverage search by multiple robots in an unknown complex terrain environment.

[0005] To achieve the above object, the present invention adopts the following technical solutions: The multi-robot full coverage search method based on the improved bionic neural network of the present invention comprises the following steps: S1, establish a grid map representing the unknown environment; each grid has the same area; S2, initializing the state of each grid, using the sensor carried by each robot to obtain the environmental information of N*N grids within the maximum coverage range of the robot, and updating the state of the N*N grids; S3, a two-dimensional bionic neural network model is established based on the grid map; each neuron corresponds to a grid, and each neuron is laterally connected to neighboring neurons; S4, initialize the neuron activity value Q. After the search starts, update the neuron activity values ​​of N*N grids within the robot's coverage area; S5, determining the next moving position of each robot according to the two-dimensional bionic neural network model and the auxiliary search mechanism.

[0006] Furthermore, the states of the grids in step S2 include unsearched grids, searched grids, and obstacle grids.

[0007] Furthermore, in step S4, the external stimulus signal corresponding to the unsearched grid neuron is defined as is the excitation signal, and the obstacle grid neuron corresponds to the external stimulus signal For inhibitory signal.

[0008] Furthermore, step S5 specifically includes: S5.1, predict the short-term optimal control input based on the robot's current position information; S5.2, construct search effectiveness function , taking the increment of neuron activity value as the main incentive and adding the turning cost; S5.3, accumulate the efficiency function of each step to obtain the L-step cumulative search efficiency function; obtain the search efficiency function values ​​of all paths within the L-step by traversing, and select the first step of the L-step prediction input corresponding to the maximum search efficiency function value As optimal control input; S5.4, establish an auxiliary search mechanism to guide the robot out of trouble when it is deadlocked. When the robot jumps out of the deadlock, continue to execute step S5.3 until the coverage search task is completed.

[0009] Furthermore, the deadlock is when the robot's state duration value Rs> deadlock threshold dead_num; the initial value of Rs is 0, when the radius is centered on the robot's current position The percentage of unsearched cells within the coverage area is less than , Rs increases by 1.

[0010] Furthermore, the guiding the robot out of trouble specifically includes temporarily expanding the robot's coverage area and planning a shortest collision-free path to the nearest unsearched grid to the current position for the robot through the A* algorithm.

[0011] Furthermore, the basis for jumping out of the deadlock is to take the current position of the robot as the center and the radius The proportion of unsearched grids within the coverage area is greater than .

[0012] The advantage of the present invention is that the improved bionic neural network model can not only help the robot to avoid obstacles effectively, but also better attract the robot to search the unknown search area. At the same time, the coverage search strategy adopts a combination of model predictive control and auxiliary search mechanism. On the basis of effectively reducing the average number of turns and average turning angle of the robot, it can help the robot get out of the local deadlock state in time, greatly improving the coverage search efficiency. The average single-step decision time for multiple robots to achieve full coverage search is in milliseconds, which can meet the needs of practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A schematic diagram of a grid map representing an unknown environment established in the method of the present invention.

[0014] Figure 2 The figure is a schematic diagram of establishing a two-dimensional bionic neural network model based on a grid map in the method of the present invention.

[0015] Figure 3 The present invention provides a flow chart for determining the next moving position of each robot according to the method of the present invention.

[0016] Figure 4 The figure is a flow chart of the auxiliary search mechanism in the method of the present invention.

[0017] Figure 5 This is a result diagram of the coverage search completed in Example 2 of the present invention.

[0018] Figure 6 This is a diagram of the neural network activity state after the search is completed in Example 2 of the present invention.

[0019] Figure 7 This is a curve showing the average coverage rate versus the number of steps using different search path planning methods. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] Example 1 The multi-robot full coverage search method based on the improved bionic neural network described in the present invention specifically comprises the following steps: S1, first establish a grid map representing the unknown environment, such as Figure 1 shown. Figure 1The five-pointed star in the middle indicates the current position of the robot, and the green dotted box indicates the robot's maximum coverage. However, the robot cannot detect the area behind obstacles. Figure 1 The black squares and red solid lines represent obstacles. The undetectable areas behind obstacles are represented in gray. The white areas represent the areas that the robot can detect. Different environmental features are represented by assigning different states to the grids. Each grid has three states: unsearched grid, searched grid, and obstacle grid. The unsearched grid corresponds to the gray area that cannot be detected behind obstacles, the searched grid corresponds to the white area that the robot can detect, and the obstacle grid corresponds to the area where the obstacles are located.

[0022] S2, at the beginning of the search, all grid states of the entire grid map are initialized to unsearched grids. Then, according to the location of the robot, the sensor carried by each robot is used to obtain the environmental information of N*N grids within the maximum coverage range of the robot, and the N*N grid map states are updated.

[0023] S3, build a two-dimensional bionic neural network model based on the grid map, such as Figure 2 shown. Figure 2 The dotted circle with a radius of r is called the receptive field of the ith neuron in neurophysiology. Neurons only respond to stimuli within their receptive fields. Each neuron has lateral connections with neurons in its neighborhood. As the connection weight, the activity values ​​between neurons are transmitted through mutual connections, and the dynamic changes of neuron activity values ​​are used to represent the dynamically changing environment. Each neuron corresponds to a grid in the grid map one by one.

[0024] S4, when the search starts, the neuron activity value Q is initialized. After the search starts, the neuron activity value is updated according to formula (1) for the N*N grids within the robot's coverage area.

[0025] Formula (1) in: Indicates The activity value of a neuron, Indicates the external stimulus signal it receives, defined as: , ; represents the stimulus input; represents inhibitory input; It is The number of neurons in the receptive field of a neuron; 𝐴, 𝐵, and 𝐷 are all positive constants, and −𝐴 represents The natural decay rate of The upper and lower limits of ; Central neuron and neighboring neurons The connection weight coefficient is defined as shown in formula (2); Representing neurons and neurons The connectivity between them is defined as shown in formula (3): Formula (2) in: Represents the state space and The Euclidean distance between them, r is the radius of the neuron’s receptive field, and 𝛼 is a positive constant; Formula (3) in: Indicates the location of neurons, For neurons The number of adjacent neurons, Representing neurons Connected neighborhood.

[0026] The changes in neuronal activity values ​​are mainly related to external stimulation signals It is related to the transmission of activity values ​​between the surrounding neighbors. Therefore, the external stimulus signal corresponding to the unsearched grid neuron is defined as is the excitation signal, which maximizes the activity value corresponding to the unsearched area, and the external stimulation signal corresponding to the obstacle grid neuron To suppress the signal, the activity value corresponding to the obstacle area is minimized, as shown in formula (4): Formula (4) in: is a sufficiently large positive constant, Indicates that the area was not searched. Indicates that the area has been searched. Indicates the obstacle area; S5, determine the next moving position of each robot according to the two-dimensional bionic neural network model and auxiliary search mechanism. Figure 3 The detailed steps are as follows: The position change of the robot can be accurately described by the kinematic model as shown in the following equation.

[0027] in: Represents a robot At the kth step, Represents a robot speed, is the time step, Indicates the direction of motion.

[0028] S5.1, according to the current position information of the robot, select As the Step control input , predict the short-term optimal control input, as shown in formula (5).

[0029] Formula (5) in: , is the number of robots, , L is the number of prediction steps. It is based on Status and control input Predict the next state of the ith robot.

[0030] S5.2, construct search effectiveness function , as shown in formula (6), in order to better guide the robot to prioritize the search of unsearched areas during the search process, the increment of neuron activity value is used as the main incentive. At the same time, in order to reduce the robot's frequent turning to reduce energy consumption, a turning cost is added.

[0031] (6) in: represents the neuron activity value increment function, which is defined as in formula (7), represents the turning cost function, which is defined as in formula (8), and represents the weight coefficient, and ; (7) in: Indicates the number of grids that the robot can cover at its current position. represents the activity value of the i-th grid, Indicates the maximum number of grids within the coverage area of ​​the current location. Represents the maximum value of neuron activity.

[0032] (8) in: represents the turning angle offset between the robot’s two-step decision, Indicates the maximum turning angle of the robot, set .

[0033] S5.3, accumulate the efficiency function of each step to obtain the L-step cumulative search efficiency function, as shown in formula (9); when the predicted next step position is an obstacle, ; (9) By traversing the search efficiency function values ​​of all paths within L steps, the first step of the L-step prediction input corresponding to the maximum search efficiency function value is selected. As the optimal control input, as shown in formula (10): (10) in: is the optimal sequence The first two-dimensional vector component of is the next moving position.

[0034] S5.4, establish an auxiliary search mechanism to guide the robot out of deadlock when it is in deadlock. When the robot jumps out of deadlock, continue to execute step S5.3 until the coverage search task is completed. The flowchart of the auxiliary search mechanism is as follows: Figure 4 shown.

[0035] At the beginning of the search, first set the deadlock threshold dead_num and the state duration value Rs, and initialize Rs=0. Then enter step S5.3, at which time the robot's working mode is called the coverage search mode. In this mode, the robot will perform a deadlock judgment every time it moves. If the robot enters a deadlock, it will be guided out of the deadlock. When it is determined that the robot is out of the deadlock, it will enter the coverage search mode again. If the search is completed, it ends.

[0036] Deadlock determination is when the robot's current position is taken as the center and the preset radius The proportion of unsearched grids within the coverage area is less than the preset threshold When Rs>dead_num, the robot is considered to be in deadlock and the escape mechanism is activated. The expression is shown in formula (11): (11) in: Indicates the status of the robot. "1" indicates that the robot is in deadlock and is in foraging mode. "0" indicates normal coverage search mode. The robot radius is The area of ​​the uncovered area within the range, is a normal number, .

[0037] The escape mechanism is to temporarily expand the robot's coverage when it is judged that the robot is in deadlock, find the unsearched grid closest to the current position as the target point, and plan the shortest collision-free path through the A* algorithm to guide the robot out of trouble.

[0038] In the process of the robot escaping from the deadlock, each step will be judged according to formula (12) to determine whether the robot has jumped out of the deadlock state. The basis for jumping out of the deadlock is based on the current position of the robot as the center and the preset radius The proportion of unsearched grids within the coverage area is greater than the preset threshold When the robot is judged to have jumped out of the deadlock state, it will switch back to the coverage search mode, that is, return to S5.3, and continue to determine the optimal path through the model predictive control method, and repeat this process until the coverage search task is completed.

[0039] (12) in: Represents a robot The current search mode, is a positive constant, , the other terms in formula (12) have the same meaning as those in formula (11).

[0040] Example 2 like Figure 5 As shown in the figure, the size of the task area to be covered is 40×40, and there are a certain number of typical obstacles distributed in the area, which are black squares in the figure, and there are two wall-type obstacles that the robot cannot cross, which are red solid lines in the figure. The number of robots is set to 4, and they can only move one grid at a time. The coverage range of each robot is set to 5*5; the parameters of the bionic neural network are set to A=B=D=0.5, , , the activity values ​​of all neurons in the neural network are initialized to 0.4; the predictive control parameters are set to , the escape parameter is set to , , .

[0041] A set of starting positions are randomly selected for the four robots, namely (10, 7), (4, 16), (3, 9) and (9, 5). The coverage search trajectory of the robots when they move 150 steps is shown in the figure below. Figure 5 As shown; Figure 5 The middle circle indicates the starting position of each robot, the five-pointed star indicates the current position of each robot, and the search path trajectory of each robot is connected by lines of different colors. The corresponding neural network activity value state diagram is as follows: Figure 6 shown.

[0042] From overlay search results Figure 5It can be seen from the figure that there is no gray uncovered grid in the task area at this time, and the trajectories generated by each robot do not collide with obstacles or pass through wall obstacles, indicating that the method proposed in the present invention completes the full coverage search task without collision. Figure 6 It can also be seen that the state of the neural network has stabilized at this time, and there are no more high activity value grids, which also indicates that all grids have been searched.

[0043] In order to further demonstrate the superiority of the present invention, the method proposed in the present invention is compared with the traditional bionic neural network method, the bionic neural network method considering connectivity and the traditional predictive control planning search path method.

[0044] The above four methods were subjected to 30 Monte Carlo experiments under the same experimental conditions, and the average coverage of the regions in the 30 experiments was calculated. The average coverage of different methods varies with the number of steps. Figure 7 As shown. Figure 7 It can be seen that the method proposed in the present invention is the best in terms of average coverage index. All 30 groups of experiments achieved full coverage, and the full coverage rate reached 100%.

Claims

1. A multi-robot full coverage search method based on an improved bionic neural network, characterized in that: The following steps are included: S1, establish a grid map representing the unknown environment; each grid has the same area; S2, initializing the state of each grid, using the sensor carried by each robot to obtain the environmental information of N*N grids within the maximum coverage range of the robot, and updating the state of the N*N grids; S3, a two-dimensional bionic neural network model is established based on the grid map; each neuron corresponds to a grid, and each neuron is laterally connected to neighboring neurons; S4, initialize the neuron activity value Q. After the search starts, update the neuron activity values ​​of N*N grids within the robot's coverage area; S5, determining the next moving position of each robot according to the two-dimensional bionic neural network model and the auxiliary search mechanism.

2. The multi-robot full coverage search method based on improved bionic neural network according to claim 1, characterized in that: The states of the grids in step S2 include unsearched grids, searched grids, and obstacle grids.

3. The multi-robot full coverage search method based on improved bionic neural network according to claim 1, characterized in that: In step S4, the external stimulus signal corresponding to the unsearched grid neurons is defined is the excitation signal, and the obstacle grid neuron corresponds to the external stimulus signal For inhibitory signal.

4. The multi-robot full coverage search method based on improved bionic neural network according to claim 1, characterized in that: Step S5 specifically includes: S5.1, predict the short-term optimal control input based on the robot's current position information; S5.2, construct search effectiveness function , taking the increment of neuron activity value as the main incentive and adding the turning cost; S5.3, accumulate the efficiency function of each step to obtain the L-step cumulative search efficiency function; obtain the search efficiency function values ​​of all paths within the L-step by traversing, and select the first step of the L-step prediction input corresponding to the maximum search efficiency function value As optimal control input; S5.4, establish an auxiliary search mechanism to guide the robot out of trouble when it is deadlocked. When the robot jumps out of the deadlock, continue to execute step S5.3 until the coverage search task is completed.

5. The multi-robot full coverage search method based on improved bionic neural network according to claim 4, characterized in that: The deadlock is when the robot's state duration value Rs> deadlock threshold dead_num; the initial value of Rs is 0, when the robot's current position is taken as the center and the radius The percentage of unsearched cells within the coverage area is less than , Rs increases by 1.

6. The multi-robot full coverage search method based on improved bionic neural network according to claim 4, characterized in that: Specifically, guiding the robot out of trouble involves temporarily expanding the robot's coverage area and planning a shortest collision-free path to the nearest unsearched grid at the current position for the robot through the A* algorithm.

7. The multi-robot full coverage search method based on improved bionic neural network according to claim 4, characterized in that: The basis for jumping out of deadlock is to take the current position of the robot as the center and the radius The proportion of unsearched grids within the coverage area is greater than .

Images