A multi-robot task planning method under uncertainty map

Abstract

The application discloses a kind of under uncertainty map multi-robot task planning method, it is related to multi-robot cooperative control technical field, can be aimed at the uncertainty of environment and planning time.The technical scheme of the application includes the following steps: analyzing unmanned aerial vehicle system, the kinematics space of unmanned aerial vehicle is obtained;Using SLAM algorithm, the working environment is mapped, to obtain the uncertainty map of target space;Offline planning algorithm is executed, for the given LTL task, multi-robot system information and prior map, the state space transition of multi-robot system is simulated by the way of building search tree, to obtain the execution plan that meets LTL constraint is solved;During task execution, online planning algorithm is executed, specifically: the information collected in real time by multi-robot system is updated to the uncertainty map, to obtain error-free landmark building information, while combining the target landmark of current execution plan, adjust the execution plan, to ensure more reliable completion of task.

Description

Technical Field

[0001] This invention relates to the field of multi-robot cooperative control technology, and specifically to a multi-robot task planning method under an uncertain map. Background Technology

[0002] With the increasing demand for multi-robot systems in fields such as intelligent manufacturing, intelligent logistics, and disaster relief, the efficient coordination of multiple robots to complete complex spatiotemporal constrained tasks has become a hot topic in invention. Traditional motion planning methods mainly focus on obstacle avoidance and single-target arrival, making it difficult to implement and describe complex tasks involving cooperation and timing, such as "periodic inspection" and "ordered task execution." Linear Temporal Logic (LTL), due to its rigorous mathematical semantics and powerful temporal expression capabilities, is widely used to describe such tasks, providing strong theoretical support for complex multi-robot collaboration.

[0003] Current multi-robot task planning algorithms mainly include traditional graph search-based algorithms and algorithms that construct search trees to search the state space. However, both graph search algorithms and search tree algorithms have some problems. Graph search algorithms experience exponential complexity increases when the number of robots increases; while search tree algorithms, because they cannot fully explore the state space, only possess asymptotic optimality and probabilistic completeness. Furthermore, current algorithms use atomic propositions to describe each task, but these atomic propositions lack constraints on probability and distance. These problems make task planning for large-scale robot systems extremely difficult. In recent years, due to the low efficiency of traditional algorithms, search tree algorithms have become a more popular choice for UAV control.

[0004] Furthermore, most current task planning algorithms are based on known, error-free deterministic maps, which is virtually impossible to implement in practice. In reality, any sensor has hardware errors and various environmental interferences, meaning the maps it constructs cannot accurately reflect the locations of buildings in the real environment. Currently, most mainstream multi-robot task planning algorithms rely on known maps and traverse the state space using search trees to find execution strategies that meet LTL constraints. However, this mainstream approach often fails to complete tasks in real-world environments due to map construction errors. This indicates that while the assumption of known, error-free maps simplifies task planning and facilitates problem solving, it cannot be the sole basis due to its potential problems. The execution plan must be revised based on environmental information collected during actual operation; the known map only serves as prior information in offline planning. For example, most current algorithms discretize continuous maps into directed graphs, where nodes represent landmarks containing geographical information. When this information is incorrect, the robot may not accurately reach the target location, leading to task failure. Therefore, fully utilizing known map information while maximizing task execution success rate and computational efficiency is crucial for multi-robot task planning problems.

[0005] Given a known prior map, how to conduct reasonable offline planning to ensure that the multi-robot system can complete the task as much as possible, and at the same time combine the sensor information collected during actual operation to ensure the reliability of task execution, is very important for task planning in uncertain environments.

[0006] In recent inventions, the feasibility of traditional graph search algorithms for known maps has been proven, but their computational efficiency is extremely low. Emerging algorithms for constructing search trees, while possessing only probabilistic completeness and asymptotic optimality, are computationally very fast and have been verified to handle task planning problems with a state space size of 9765625, which is beyond the capabilities of traditional graph search algorithms. Following this, the synthesis of time-logic control with uncertainty has received increasing attention in inventions, proposing methods relying on probabilistic model checking to synthesize controllers that maximize probabilities and considering the uncertainty of incomplete (or dynamic) environment models in the workspace. Y. Kantaros and GJ Pappas then proposed an abstraction-free method for time-logic planning in uncertain environments, modeled using probabilistic semantic maps obtained from SLAM algorithms, expanding atomic propositions, and introducing probabilistic and distance constraints, effectively addressing the problem of map uncertainty.

[0007] Most existing multi-robot task planning methods are based on the assumption of known and error-free deterministic maps, lacking effective modeling and utilization of environmental uncertainties. In practical applications, due to sensor noise, environmental occlusion, and mapping errors, there are often deviations between the prior map and the real environment, making it difficult to accurately execute the execution strategy obtained based on offline planning, and compromising the reliability of task completion. Furthermore, existing methods typically separate offline planning from online execution, lacking a unified mechanism for updating the environmental model and dynamically correcting existing plans by incorporating real-time perception information during task execution. This results in insufficient adaptability and robustness of the system in complex and uncertain environments. In addition, for multi-robot task planning problems satisfying linear temporal logic constraints, traditional graph search-based methods experience a sharp increase in computational complexity as the task size increases. While search tree-based planning methods are computationally efficient, their sampling strategies lack effective guidance on temporal logic objectives and uncertainty information, easily generating a large number of invalid samples, leading to long planning times and making it difficult to simultaneously meet the reliability and real-time requirements of task execution.

[0008] Currently, commonly used multi-robot task planning algorithms mainly employ search tree construction. While this algorithm offers higher efficiency compared to traditional graph search algorithms, it cannot guarantee the optimal solution. It is further divided into centralized, distributed, and hybrid algorithms. Centralized planning algorithms are theoretically rigorous and can guarantee global optimality, but suffer from poor real-time performance and weak adaptability to dynamic environments. Distributed planning algorithms are computationally efficient but cannot guarantee optimality and exhibit weak global task collaboration. Hybrid planning algorithms balance global optimality and real-time performance, making them suitable for large-scale dynamic scenarios, but their architecture is complex.

[0009] It is evident that existing task planning methods cannot achieve both reliable task execution and good real-time performance. Summary of the Invention

[0010] In view of this, the present invention provides a multi-robot task planning method under uncertainty map, which can perform multi-robot task planning to address the key issues of robot task planning—environmental uncertainty and planning time, and has good real-time performance while ensuring the reliability of task execution.

[0011] To achieve the above objectives, the technical solution of the present invention includes the following steps:

[0012] The first step is to analyze the UAV system. Based on the UAV's parameters and the inner loop of the flight controller's angular velocity, the kinematic space of the UAV is solved. The second step is to use the SLAM algorithm to build a map of the working environment, thereby obtaining an uncertainty map of the target space; The third step is to execute an offline planning algorithm. Given the LTL task, multi-robot system information, and prior map, the algorithm simulates the transition of the multi-robot system in the state space by constructing a search tree, thereby solving for an execution plan that meets the LTL constraints. During task execution, an online planning algorithm is executed, specifically: the uncertainty map is updated by combining information collected in real time by the multi-robot system to obtain error-free landmark building information, and the execution plan is adjusted by combining the target landmarks of the current execution plan to ensure more reliable task completion.

[0013] Furthermore, the SLAM algorithm is used to construct a map of the working environment, resulting in a probabilistic semantic map of the workspace. The specific steps are as follows: A probabilistic semantic map constructed using the SLAM algorithm contains the locations of various landmarks and their uncertainty distributions. The uncertainty of the landmark locations is represented by a Gaussian distribution. It means that among them Indicates landmarks in a known map Location, Indicates landmarks in a known map The covariance matrix is ​​used to represent the landmark location. Uncertainty; Robots with Operating in an unknown environment with several landmarks, the semantic map of the environment is defined as a set of... There are n elements, each element is represented as... ;in It is a Gaussian random variable, representing a value with mean. Covariance Matrix The location of the landmark, and It is a discrete random variable representing the class of landmarks; ,in It is the number of different classes that exist in the environment.

[0014] Furthermore, the control inputs for each robot include angles. and distance traveled , where j is the j-th robot.

[0015] Furthermore, the output of the offline planning algorithm is the global path. .

[0016] Furthermore, the output of the online planning algorithm is the adjusted local path. .

[0017] Furthermore, the offline planning algorithm has the following specific steps: With the maximum number of iterations Robot dynamics model, prior map Initial position and the given LTL task As input, the LTL task is transformed into the corresponding Büchi automaton NBA; and the node set of the search tree is initialized. Edge set Subset partitioning Cost function and target node set ; The number of iterations has not exceeded the upper limit. Under the given conditions, according to the given probability density function By sampling a subset of nodes, a sample set is obtained. and the sampling nodes therein Perform cyclic processing; For each sampling node, further sampling control input is performed. Calculate the new position state and corresponding NBA status If the new state meets the feasibility conditions, then the new node will be... Add to the search tree and update the node set. Edge set The system calculates the cost function J and determines whether a new NBA state exists. If the NBA state corresponding to the new node is the receiving state, the new node is added to the target node set. ; When the iteration ends or the maximum number of iterations is reached, the target node set is then... The search tree is used to find feasible paths from the initial node to the target node. An optimization problem is constructed by minimizing the cost function and the optimal path is obtained by solving it. The final output of the offline planning algorithm is the global path.

[0018] Furthermore, by minimizing the cost function, an optimization problem is constructed and solved to obtain the optimal path. Specifically: For a given global LTL specification and the control signal set of each robot and uncertainty maps A global path needs to be determined. Make , Indicates that the specification is met, stopping time and a control input sequence Furthermore, minimizing the path length, the optimization problem can be expressed as: (1) in, Cost function Indicates the stop time. This represents a control input sequence. Indicating a multi-robot system Location at any given moment Indicates the position from time t Position at time t+1 The price paid At the position of the dynamic equation Control input is The value corresponding to the time; This represents the global LTL specification that needs to be satisfied. This represents the set of control signals for each robot. Indicates global path Meets LTL specifications .

[0019] Beneficial effects: This invention provides a planning method that ensures both the reliability of task execution and the efficiency of task planning. First, it obtains the execution strategy of a multi-robot system through offline planning. Then, combining image information of the surrounding environment and the drone's position information acquired by the drone, an online planner updates the uncertain map and makes local corrections to the offline execution plan, improving the reliability of task completion. This method can adjust the execution plan through online planning when there is a discrepancy between the known map and the actual map, further improving the reliability of task execution. Compared to previous task planning methods, this invention ensures both reliability and good real-time performance. Attached Figure Description

[0020] Figure 1 The uncertainty map involved in this invention Figure 2 This is a flowchart illustrating the offline planning algorithm of the present invention.

[0021] Figure 3 This is a flowchart illustrating the sampling algorithm in this invention; Figure 4 This is a schematic diagram of the online planning process of the present invention. Detailed Implementation

[0022] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0023] This invention provides a planning method that ensures both the reliability of task execution and the efficiency of task planning. The method includes the following steps: The first step is to analyze the UAV system. Based on the UAV's mass, size, and other parameters, and combined with the inner loop of the flight controller's angular velocity, the UAV's kinematic space is solved.

[0024]

[0025] in This represents the three-dimensional position coordinates (in meters) of the UAV in the world coordinate system at time k. This represents the roll, pitch, and yaw angles at time k (unit: rad). This represents the linear velocity component of the UAV in the world coordinate system at time k (unit: m / s). These represent roll / pitch / yaw angular velocities (unit: rad / s). The second step is to use the SLAM algorithm to map the working environment, thereby obtaining an uncertainty map of the target space.

[0026] The third step involves offline planning for a given LTL task based on a known uncertainty map. A cost function is constructed, and the state transitions of the multi-robot system in the workspace are simulated using a search tree. The execution strategy that minimizes the objective cost is then determined. Simultaneously, the uncertainty map is updated based on the positions of each robot and the sensor ranges during the planning process, providing information for subsequent planning. During execution, the online planner incorporates the collected information to revise the current plan, ensuring more reliable task completion.

[0027] The algorithm of this invention is applicable to Ubuntu 20.04, written in Python 3.8, and uses ROS as the software tool for scheduling various tasks. The algorithm includes: The offline planning algorithm, given an LTL task, multi-robot system information, and prior map, simulates the transitions of the multi-robot system in the state space by constructing a search tree, thereby solving for an execution plan that meets the LTL constraints.

[0028] The online planning algorithm updates the uncertain map by combining information collected in real time by the multi-robot system during task execution, obtaining error-free landmark building information. At the same time, it adjusts the execution plan based on the target landmarks of the current execution plan to ensure more reliable task completion.

[0029] Uncertainty maps such as Figure 1 As shown in the figure, the location of each landmark building and its uncertainty distribution are included. The uncertainty of the landmark location is expressed using a Gaussian distribution. It means that, among them Indicates landmarks in a known map Location; Indicates landmarks in a known map The covariance matrix is ​​used to represent the landmark location. Uncertainty.

[0030] This invention assumes that the robot has The map operates in an unknown environment with several landmarks. The uncertainty of the environment is defined as a set of... There are n elements, each element is represented as

[0031] in It is a Gaussian random variable, representing a value with mean. Covariance Matrix The location of the landmark, and It is a discrete random variable representing the class of the landmark. This study assumes... ,in It is the number of different classes that exist in the environment.

[0032] For a given global LTL specification and the control signal set of each robot and uncertainty maps A global path needs to be determined. Make , Indicates that the specification is met, stopping time and a control input sequence Furthermore, minimizing the path length transforms it into an optimization problem, namely... (1) Cost function Indicates the position from time t Position at time t+1 The price paid At the position of the dynamic equation Control input is The corresponding value at time (note that it corresponds to the dynamic equation in the first step). The offline planning algorithm uses an improved RRT* sampling algorithm to construct the search tree. The method simulates the robot's transitions in the state space by constructing a search tree, which simultaneously explores the motion space of the multi-robot system and the corresponding LTL formula. The state space of the uncertain Büchi automaton. Represents the set of nodes of a tree. Represents the set of edges of the search tree. Represents the distance from the root node to the node in the search tree. The movement distance, where node information includes the position information, angle information, Büchi state, and cost of the current node for each robot. The nodes... The cost is calculated as (2) Because of the introduction of uncertainty, and because the transitions between macroscopic states of the robot cannot well meet the requirements of probabilistic tasks, the planning results need to be downgraded from state sequences to position coordinates during task planning. Downgrading to position coordinates can also avoid the complex path cost calculation problem caused by obstacles between paths during robot state transitions. At the same time, it enables multi-robot task planning in obstacle-filled environments, directly obtaining the global path points of each robot, avoiding the need to use path planning algorithms to calculate robot transition paths after obtaining the robot state sequence.

[0033] In this embodiment of the invention, the offline planning algorithm process is as follows: Figure 2 As shown, specifically: First, consider the maximum number of iterations. Robot dynamics model, prior map Initial position and the given LTL task As input, the LTL task is transformed into the corresponding Büchi automaton (NBA), and the node set of the search tree is initialized. Edge set Subset partitioning Cost function and target node set If the number of iterations does not exceed the upper limit. Under the given conditions, the algorithm follows the given probability density function. By sampling a subset of nodes, a sample set is obtained. and the sampling nodes therein Perform loop processing; for each sampling node, further sample the control input. Calculate the new position state and corresponding NBA status If the new state meets the feasibility conditions, then the new node will be... Add to the search tree and update the node set. Edge set The system calculates the cost function J and determines whether a new NBA state exists. If the NBA state corresponding to the new node is a receiving state, it adds the node to the target node set. When the iteration ends or the maximum number of iterations is reached, the algorithm proceeds with the next iteration from the target node set. The search tree is used to find feasible paths from the initial node to the target node, and the optimal path is obtained by minimizing the cost function. This will be the final output.

[0034] This invention will use the LTL formula Transform it into an undetermined Büchi automaton (NBA). Then initialize the search tree. Initialized to , , and Furthermore, the search tree is used in each iteration of the offline planning algorithm. In China, through... Add a new status He Xiang Add the corresponding edges to incrementally expand the search tree. Then... The update is performed for the next iteration, where k represents the k-th iteration. After... After the nth iteration (where (You can configure it according to your needs). Once the search tree is built, it is finally determined based on its target node set. And search the search tree to find the global path that minimizes cost (1). Furthermore, in the process of constructing the search tree, this invention simultaneously considers the detection range and noise of the robot's sensors, simulates the interaction between the robot and the actual environment, and updates the uncertainty map using Kalman filtering in conjunction with the current robot position information when a landmark appears within the sensor range.

[0035] For sets The steps to update are as follows: If a status update already exists, the update will be performed as follows: (New NBA Node) Related Subset ,but Otherwise, create a new collection. ,Right now (Explanation follows), and .

[0036] However, the NBA data obtained through direct conversion contains many redundant states and conversions, and transfer relationships. Many transition conditions in the set cannot be achieved in the current multi-robot system. For example, the number of atomic propositions that need to be satisfied is greater than the number of robots. In order to simplify subsequent calculations, avoid the generation of useless nodes, and speed up the construction of the search tree, these states need to be pruned.

[0037] This invention first trims the NBA by removing state transition edges that are impossible to achieve, which would require a robot. NBA transitions that simultaneously satisfy atomic propositions for multiple positions are impossible to achieve. For example, if an NBA transition occurs in... It is valid when it is true, and the region and If the regions are disjoint, the transformation is classified as infeasible; otherwise, it is feasible. Therefore, to simplify the problem, this invention assumes that all regions... They are all non-intersecting.

[0038] Secondly, this invention defines a distance function. , This represents the NBA state set, used to measure the state of the NBA after the trimming process. to state The minimum number of conversions required. This function is defined as follows: (3) in Indicates from the trimmed NBA arrive The shortest path, This indicates the cost.

[0039] This invention addresses the issue of long planning times in current algorithms for complex problems. It improves upon the sampling algorithm by introducing dynamic distance, accelerating the construction of the search tree and thus increasing the efficiency of solving the execution strategy. The planned path input from the offline controller, combined with real-time acquired information, allows for rapid adjustments to the execution plan, enabling the UAV to accurately complete the given task.

[0040] In this embodiment of the invention, the Sample algorithm in the offline planning algorithm is the sampling algorithm.

[0041] In each iteration of the offline planning algorithm In the middle, a new state Generated by a sampling algorithm. State The construction consists of two steps: sampling a location state. Then And an NBA status Fusion, thus obtaining Finally, add it to the tree structure.

[0042] In order to construct state This invention first sets the node set according to NBA state components. Divide into a finite number of subsets, denoted as ,Right now It contains a set of nodes. All NBA statuses in China are Tree nodes. Through The construction of this invention yields ,in It is an algorithm During the next iteration The number of subsets. Furthermore, note that due to the structure of the NBA, It is finite. In the offline planning algorithm... During the next iteration, i.e., initialization, we have , .

[0043] Secondly, given a set This invention first starts with discrete probability distribution Sampling an index Density function The offline planning algorithm is defined in the first... In the next iteration, from the node set Select collection The probability of any density function. Both can be used for sample extraction. As long as it satisfies the following assumptions: Assumption 1 (probability density function) ): (i) Probability density function satisfy And for all For a certain ,Should In all iterations (ii) Independent samples From Extract from.

[0044] This assumption guarantees that all control inputs in the control input space have a certain sampling probability (with a lower bound) and that each sampling is independent, laying the foundation for the algorithm's effectiveness and theoretical analysis. It ensures that the algorithm does not miss any states, avoids getting trapped in local optima, and helps explore the entire control space.

[0045] Next, given from medium-sampled set And for each state The present invention originates from a discrete distribution Sampling a control input Given from Control input for intermediate sampling Then, the construction status of the present invention for Add it to the tree. Any density function. Both can be used for sample extraction. As long as it satisfies the following assumptions: Assumption 2 (probability density function) ): (i) Distribution function satisfy For all For a certain ,Should In all iterations (ii) Independent samples From the probability density function Extract from.

[0046] Similar to Assumption 1, Assumption 2 is also to ensure that the algorithm does not ignore any possibilities, avoids getting trapped in local optima, and guarantees the independence of sampling in each iteration.

[0047] Clearly, the discrete uniform distribution meets the requirements for both assumptions 1 and 2, but... and The choice of [function name] affects the algorithm's performance. This invention employs two probability density functions with guiding effects to facilitate the rapid generation of execution strategies. Details of the invention are as follows: probability density function First, this invention defines a set of... This set collects all nodes. These nodes are All nodes have the target state minimum distance , recorded as ,Right now

[0048] gather Initially, it only contains the root node, and it is updated as new state is added to the tree (and...). Given a set This invention defines a set of... This set collects those that satisfy subset of index Given a set probability density function Defined as: (4) in Indicates selecting any that satisfy subset of The probability of . Among them, It can vary with the number of iterations. Dynamically changing, but it should always satisfy This is to ensure a closer bias towards the receiving state.

[0049] probability density function : Sampling function The invention enabled the connection in the NBA through the trimmed process. arrive The shortest path (in steps) to reach the state. First, given a state This invention calculates the next NBA state that the current state can transition to.

[0050] Next, we construct the reachable set. This set collects all that can be obtained from Achieve NBA status in one step in the trimmed NBA. Defined as (5) exist Of all states, the present invention selects and The state with the minimum distance is denoted as .

[0051] To achieve from arrive For the transfer to occur, the robotic system needs to meet its transfer conditions. The present invention then guides the generation of corresponding position coordinates based on this condition. Therefore, the present invention first selects the present Neutralizing and integrating robots The relevant atomic propositions, and then the present invention can obtain a robot. Atomic propositions that need to be satisfied For example, corresponding to the label atomic propositions yes Furthermore, if for all , The computational robot of this invention It can make All positions that are true. For example, if Then all those with non-zero probability and belonging to Landmarks are all collected in the collection In. Of all regions, the present invention selects one region, denoted as... This area is selected to choose a new control input. This makes the robot Orientation Movement. Next, this invention discusses how to select... For robots If the atomic proposition The form is ,but ,therefore It enables robots Towards The control signal for motion. If the atomic proposition... The form is The present invention is from Select landmark Make the value The largest, of which It is the covariance matrix The determinant of the matrix, value The larger the location and landmarks The less uncertainty there is in the category, the higher the reliability of completing the task. For robots Then let .

[0052] Given assigned to the robot area This invention constructs a density function This invention relates to robots control input set The sampling probabilities are as follows: (6) in Related to the previous text In Similar, both must satisfy To quickly reach the target point, this invention optimizes the control signal by introducing dynamic distance and optimal control input. No longer fixed but moving with the robot To the target area The distance changes, and the distance traveled is calculated as follows: (7) in For robots Point to target area Angle, Indicates the shortest distance to travel. Represents robots Distance between current location and the center of target area You can specify it yourself. .

[0053] The overall process of the sampling algorithm is as follows: Figure 3 exhibit.

[0054] After the construction of the new node is completed, the present invention will use the Kalman filter algorithm to update the location information of landmarks in the uncertainty map that appear within the robot's perception range according to the current location of the robot system. This will simulate the operation of the robot system in a real environment, provide more information for the construction of the subsequent search tree, and make the algorithm more reasonable.

[0055] When executing offline plans on an uncertain map, this invention aims to allow the robot to update the map based on current sensor data and modify its execution plan. To this end, this invention provides an online planner, the overall process of which is as follows: Figure 4 As shown, this module is designed for drones equipped with depth cameras. During drone operation, the depth camera captures image information below the drone. To simplify the problem, this invention abstracts the drone's field of view into a rectangular area, the range of which is related to the drone's height. When a landmark appears within the drone's field of view, the algorithm calculates the actual position of the landmark based on the drone's current pose and image information, and updates the landmark's Gaussian distribution information using Kalman filtering. The position calculation algorithm is related to robot vision and will not be elaborated upon here.

[0056] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-robot task planning method under uncertain maps, characterized in that, include: The first step is to analyze the UAV system. Based on the UAV's parameters and the inner loop of the flight controller's angular velocity, the kinematic space of the UAV is solved. The second step is to use the SLAM algorithm to build a map of the working environment, thereby obtaining an uncertainty map of the target space; The third step is to execute an offline planning algorithm. Given the LTL task, multi-robot system information, and prior map, the algorithm simulates the transition of the multi-robot system in the state space by constructing a search tree, thereby solving for an execution plan that meets the LTL constraints. During task execution, an online planning algorithm is executed, specifically: the uncertainty map is updated by combining information collected in real time by the multi-robot system to obtain error-free landmark building information, and the execution plan is adjusted by combining the target landmarks of the current execution plan to ensure more reliable task completion.

2. The multi-robot task planning method under an uncertain map as described in claim 1, characterized in that, The SLAM algorithm is used to construct a map of the working environment, resulting in a probabilistic semantic map of the working space. The specific steps are as follows: A probabilistic semantic map constructed using the SLAM algorithm contains the locations of various landmarks and their uncertainty distributions. The uncertainty of the landmark locations is represented by a Gaussian distribution. It means that, among them Indicates landmarks in a known map Location, Indicates landmarks in a known map The covariance matrix is ​​used to represent the landmark location. Uncertainty; Robots with Operating in an unknown environment with several landmarks, the semantic map of the environment is defined as a set of... There are n elements, each element is represented as ;in It is a Gaussian random variable, representing a value with mean. Covariance Matrix The location of the landmark, and It is a discrete random variable representing the class of landmarks; ,in It is the number of different classes that exist in the environment.

3. The multi-robot task planning method under an uncertain map as described in claim 1, characterized in that, The control inputs for each robot include angles. and distance traveled , where j is the j-th robot.

4. The multi-robot task planning method under an uncertain map as described in claim 1, characterized in that, The output of the offline planning algorithm is the global path. .

5. The multi-robot task planning method under an uncertain map as described in claim 1, characterized in that, The output of the online planning algorithm is the adjusted local path. .

6. The multi-robot task planning method under an uncertain map as described in claim 1, characterized in that, The offline planning algorithm has the following specific steps: With the maximum number of iterations Robot dynamics model, prior map Initial position and the given LTL task As input, the LTL task is transformed into the corresponding Büchi automaton NBA; and the node set of the search tree is initialized. Edge set Subset partitioning Cost function and target node set ; The number of iterations has not exceeded the upper limit. Under the given conditions, according to the given probability density function By sampling a subset of nodes, a sample set is obtained. and the sampling nodes therein Perform cyclic processing; For each sampling node, further sampling control input is performed. Calculate the new position state and corresponding NBA status If the new state meets the feasibility conditions, then the new node will be... Add to the search tree and update the node set. Edge set The system calculates the cost function J and determines whether a new NBA state exists. If the NBA state corresponding to the new node is the receiving state, the new node is added to the target node set. ; When the iteration ends or the maximum number of iterations is reached, the target node set is then... The search tree is used to find feasible paths from the initial node to the target node. An optimization problem is constructed by minimizing the cost function and the optimal path is obtained by solving it. The final output of the offline planning algorithm is the global path.

7. The multi-robot task planning method under an uncertain map as described in claim 6, characterized in that, The optimal path is obtained by minimizing the cost function to construct an optimization problem and solving it. Specifically: For a given global LTL specification and the control signal set of each robot and uncertainty maps A global path needs to be determined. Make , Indicates that the specification is met, stopping time and a control input sequence Furthermore, minimizing the path length, the optimization problem can be expressed as: （1） in, Cost function Indicates the stop time. This represents a control input sequence. Indicating a multi-robot system Location at any given moment Indicates the position from time t Position at time t+1 The price paid At the position of the dynamic equation Control input is The value corresponding to the time; This represents the global LTL specification that needs to be satisfied. This represents the set of control signals for each robot. Indicates global path Meets LTL specifications .

Images