A multi-robot autonomous exploration method and system

Abstract

The present invention discloses a multi-robot autonomous exploration method and system, the system includes a single robot platform, each platform has a differential wheel robot chassis, a 16-line 3D mechanical laser radar, an industrial computer, and a router; each industrial computer has two network ports, the laser radar and the router respectively transmit data with the industrial computer through the corresponding network ports, and the two are set in different network segments; the robot chassis is connected to the industrial computer through a CAN bus for data transmission. Select multiple routers that can perform Mesh networking, set the main router, and other routers are added to the local area network of the main router as sub-routes. The present invention relates to a graph-based multi-robot autonomous exploration algorithm, which efficiently covers the environment by constructing a local uniform undirected graph and a global random undirected graph for each car, and finally realizes the collaborative exploration of unknown space by a multi-robot system, which can achieve relatively fast and accurate exploration, and the exploration trajectory of each robot is relatively smooth.

Description

Technical Field

[0001] The present invention relates to a multi-robot autonomous exploration method and system, belonging to the technical field of mobile robot autonomous exploration. Background Art

[0002] Mobile robotics, as a technology that can replace human labor, offers numerous advantages. They can replace humans in repetitive, mechanical tasks, as well as dangerous tasks. With the development of numerous related technologies for mobile robots, the potential for significant benefits, and the safety benefits they offer, mobile robotics are increasingly evolving towards multi-machine systems, expanding their scope of application.

[0003] The Simultaneous Localization and Mapping (SLAM) method is a currently more commonly used synchronous map construction and positioning method. It can construct maps more accurately and over a large area, and can include more details. For large-scale environments and scenes with many narrow and long channels, the efficiency of single-robot SLAM is ultimately limited. Therefore, it is very necessary to study multi-robot collaborative SLAM methods in multi-robot systems to improve map exploration efficiency.

[0004] Autonomous exploration by mobile robots aims to enable them to rapidly, completely, and accurately observe unknown areas within an unknown, bounded environment, thereby building a complete map of the environment for navigation and planning. The solution involves continuously searching for a series of observation points near the boundaries of the unknown area and controlling the robots to visit these points, ultimately completing the exploration of the unknown area. The efficiency of autonomous exploration by a single robot is relatively limited, necessitating the development of multi-robot autonomous exploration systems and related algorithms. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a multi-robot autonomous exploration method and system to realize the collaborative exploration of unknown space by a multi-robot system, which can achieve relatively fast and accurate exploration, and the exploration trajectory of each robot is relatively smooth.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] A multi-robot autonomous exploration method comprises the following steps:

[0008] Step 1: Given the initial position of each robot, initialize each robot. Each robot moves forward six meters to construct a preliminary environment around the robot's initial position.

[0009] Step 2: Determine whether the current number of iterations has reached the maximum number of iterations or whether a stop instruction has been received. If so, the iteration ends; otherwise, the exploration process continues.

[0010] Step 3: Use the multi-machine SLAM algorithm to obtain the laser point cloud information and positioning information of each robot, and construct the TSDF map for local and global exploration;

[0011] Step 4: For each robot, expand the global graph according to the robot trajectory, randomly sample global graph nodes around the robot and perform screening and expansion;

[0012] Step 5: uniformly sample the local graph nodes around the robot to construct a local graph and screen obstacles, and output the local graph and its status;

[0013] Step 6: Determine whether the local graph state output in step 5 is successful. If so, calculate the voxel gain for all leaf nodes in the local graph, mark the leaf nodes with voxel gain greater than a first preset threshold as local graph boundary nodes, and output the local graph and its state. Otherwise, proceed to step 9 to perform global exploration.

[0014] Step 7: Determine whether the local graph state output in step 6 is successful. If so, solve the traveling salesman problem for the local graph boundary nodes to obtain the shortest visit sequence of the local graph boundary nodes and output the local graph state. Otherwise, go to step 9 to perform global exploration.

[0015] Step 8: Determine whether the local graph status output in step 7 is successful. If so, concatenate the paths according to the shortest access order and perform path optimization. Use the local graph boundary nodes in the path as nodes of the global graph, and use the path formed by connecting two adjacent local graph boundary nodes in the path as the edge of the global graph. Output the final path of the local exploration. Otherwise, proceed to step 9 to perform global exploration.

[0016] Step 9: Calculate the exploration gain and voxel gain of all nodes in the global graph to obtain the boundary nodes of the global graph, and obtain the global graphs of other robots for fusion;

[0017] Step 10: For each robot, use the AStar algorithm to calculate the shortest path for the robot to reach the nearest global graph boundary node;

[0018] Step 11: The number of iterations is increased by one. The final path obtained by local exploration of each robot or the shortest path obtained by global exploration is optimized to obtain a tracking path. The tracking path is returned to each robot for tracking, and then the process returns to step 2.

[0019] As a preferred embodiment of the method of the present invention, in step 11, before optimization, a multi-path conflict detection algorithm is used to detect whether there is a conflict between the final path or the shortest path explored by each robot, that is, the priority of each robot is pre-set, and the final path or the shortest path explored by each robot is divided into short paths. The expansion of each robot is used to detect whether there is an overlap between the short path divided by one robot and the segment path divided by another robot. If there is an overlap, the short path corresponding to the robot with the higher priority level is selected according to the priority.

[0020] As a preferred solution of the method of the present invention, in step 1, the multi-machine SLAM algorithm is used to obtain the laser point cloud information and positioning information of each robot, and the TSDF map is constructed. The specific process is as follows:

[0021] (1) Perform point cloud preprocessing on each robot, delete the robot's own point cloud within a set distance around the robot, sample the point cloud at a fixed resolution and store it;

[0022] (2) Performing stand-alone pose calculation for each robot, including instantaneous inter-frame pose estimation, key frame selection, and frame image matching;

[0023] (3) Subscribe to the point cloud data currently stored by each robot, perform point cloud downsampling, and then accumulate the current point cloud of each robot, the downsampled point cloud, and the estimated pose to obtain the current global constructed map point cloud.

[0024] As a preferred solution of the method of the present invention, the instantaneous inter-frame pose estimation is performed by comparing the point cloud of the current frame And the stored point cloud of the previous frame Get the six-degree-of-freedom transformation T in the robot sensor coordinate system i L ; The inter-frame coordinate transformation is obtained by minimizing the error of the objective function. The objective function is as follows:

[0025]

[0026] in, Represents the current frame point cloud P i The 1st to nth points in Represents the point cloud P of the previous frame i-1 The 1st to nth points in e j represents the directed distance between the jth pair of points after coordinate transformation, Indicates a point The covariance matrix of Indicates a point The covariance matrix of , the upper right corner T represents the matrix transpose.

[0027] As a preferred solution of the method of the present invention, the key frame is selected by defining a key frame structure, which includes a frame of point cloud and the current posture of the robot. After obtaining the current frame point cloud, the current posture estimate of the robot is obtained by calculation. If the current posture estimate exceeds the second preset threshold value compared with the posture transformation of the previous key frame, the current frame is selected as the key frame and added to the key frame queue.

[0028] As a preferred solution of the method of the present invention, the frame map matching is performed by matching the current frame point cloud with the current local submap, traversing and matching the current frame with each key frame in the key frame queue, and finding the distance pose after the traversal matching. The latest k key frames update the local submap, and the frame map matching solves the pose transformation by minimizing the objective function error. The objective function is as follows:

[0029]

[0030] Among them, T i L represents the transformation of six degrees of freedom in the robot sensor coordinate system, Represents the global coordinate transformation calculated in the previous frame in the world coordinate system, T i W Represents the global coordinate transformation calculated for the current frame in the world coordinate system, e j represents the directed distance between the jth pair of points after coordinate transformation, Indicates a point The covariance matrix of Represents the frame s in the local submap i j The covariance matrix of , the upper right corner T represents the matrix transpose.

[0031] A multi-robot autonomous exploration system, comprising: multiple robot platforms, each robot platform comprising a robot chassis with differential wheels, a 3D mechanical laser radar, an industrial computer, and an onboard router, wherein the 3D mechanical laser radar, the industrial computer, and the onboard router are all arranged on the robot chassis; the industrial computer has two network ports, the 3D mechanical laser radar transmits data to the industrial computer via one of the network ports, and the onboard router transmits data to the industrial computer via the other network port, the two network ports being arranged in different network segments, and the robot chassis is connected to the industrial computer via a CAN bus for data transmission; each robot platform corresponds to an ad hoc network router, and each router is self-organized via Mesh; a central control terminal, the ad hoc network router, and the industrial computer and onboard router of the robot platform form a local area network;

[0032] The 3D mechanical laser radar is used to obtain laser point cloud information around the robot;

[0033] The industrial computer is used to deploy the multi-robot autonomous exploration algorithm, receive control instructions from the central control terminal, and control the robot platform to obtain point cloud map information, posture information, local and global undirected graphs constructed during the robot platform's exploration process, and the robot's historical trajectory;

[0034] The onboard routing is used to exchange information with other robot platforms and the central control terminal;

[0035] The ad hoc network router is used to expand the local area network to ensure the effectiveness of communication between each on-board router and the central control terminal;

[0036] The central control terminal is used to send control instructions to the industrial control computer of each robot platform and send the short path to each robot platform after path conflict detection and segmentation. At the same time, the central control terminal receives point cloud map information, posture information, local and global undirected graphs constructed during the exploration process of each robot platform, and historical trajectory information of each robot.

[0037] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0038] 1. The present invention constructs a multi-robot autonomous exploration system, builds multiple vehicle parts and a multi-vehicle communication system, and can achieve multi-machine communication interaction and multi-machine collaborative work in a large-scale environment when deployed in actual machines.

[0039] 2. Based on the multi-robot system, the present invention extends the single SLAM method to the multi-robot system, realizes multi-robot collaborative mapping and positioning after initial pose calibration, and can achieve accurate positioning of each machine in the multi-robot system and collaborative map construction.

[0040] 3. The present invention develops a graph-based multi-robot autonomous exploration algorithm, which efficiently covers the environment by constructing a local uniform undirected graph and a global random undirected graph for each car, and ultimately enables the multi-robot system to collaboratively explore unknown spaces. It can achieve relatively fast and accurate exploration, and the exploration trajectory of each robot is relatively smooth. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a system overall hierarchical architecture diagram of a multi-robot autonomous exploration system and method;

[0042] Figure 2 It is an architectural diagram of the multi-robot autonomous exploration system of the present invention;

[0043] Figure 3 This is the experimental result of two robots collaboratively building a map in a ground experiment;

[0044] Figure 4This is the experimental result of three robots collaboratively constructing a map in a ground experiment;

[0045] Figure 5 is a flow chart of the multi-robot autonomous exploration method of the present invention;

[0046] Figure 6 is the computational model of node voxel gain;

[0047] Figure 7 is the path after the multi-robot path conflict detection algorithm;

[0048] Figures 8-11 It is a simulation process of multi-machine exploration;

[0049] Figure 12 It is the complete point cloud map obtained at the end of exploration. DETAILED DESCRIPTION

[0050] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be interpreted as limiting the present invention.

[0051] The overall hierarchical architecture of the multi-robot autonomous exploration system and method of the present invention is as follows: Figure 1 As shown in the figure, it mainly includes the UI user layer, communication layer, logic algorithm layer, driver layer and hardware layer. The UI user layer mainly includes human-computer interaction buttons, planning management interface and information visualization; the communication layer is a Mesh self-organizing local area network, which transmits information between the host computers of each robot and the information of interaction between the host computers of each robot and the central control host through the local area network; the logic algorithm layer mainly includes the implementation of the multi-robot autonomous exploration system algorithm, including the implementation of the multi-machine SLAM algorithm and the graph-based multi-robot autonomous exploration algorithm; the driver layer is mainly the ROS interface driver of the car chassis and the lidar sensor; the hardware layer includes the actual car chassis, sensors and communication equipment. Next, it will be introduced from the following two aspects:

[0052] 1. System Architecture

[0053] Build a multi-robot autonomous exploration system, the system architecture is as follows Figure 2As shown, a single robot platform is first constructed. Each platform consists of a differential-wheeled robot chassis, a 16-line 3D mechanical laser radar, an industrial computer, a router, and other necessary hardware. Each industrial computer has two network ports, with the laser radar and the industrial computer communicating via the network ports, and the router and the industrial computer also communicating via the network ports. The two are located on different network segments. The robot chassis connects to the industrial computer via the CAN bus for data transmission. A multi-robot communication system is then constructed, using multiple routers capable of mesh networking. A master router is designated, and other routers are added to the master router's local area network as sub-routers. Finally, a human-machine interface, planning controller, and visualization interface are constructed on the central control terminal, enabling communication between terminals via the local area network.

[0054] 2. Exploration Methods

[0055] (1) Multi-machine SLAM algorithm

[0056] Before deploying the autonomous exploration algorithm, it is necessary to incrementally construct a point cloud map based on the point cloud information of each machine and perform positioning of each machine. The specific implementation steps of the multi-machine SLAM algorithm in the present invention are as follows:

[0057] First, point cloud preprocessing is performed to delete the robot's own point cloud within a set distance around the current robot, and the point cloud is sampled and stored at a fixed resolution.

[0058] Secondly, the single-machine pose calculation is performed, including instantaneous inter-frame pose estimation, key frame selection, and frame image matching; point cloud P and point p are defined. i ∈P, all located in the world R in the Cartesian coordinate system 3 In it, L is the lidar coordinate system, W is the world coordinate system, S is the subgraph, C is the covariance, and T is the pose.

[0059] The instantaneous inter-frame pose estimation is performed by comparing the point clouds of this frame And the stored previous frame point cloud information Get the coordinate transformation T in the robot sensor coordinate system i L , and solve the inter-frame matching objective function to obtain the inter-frame coordinate transformation; the objective function is as follows:

[0060]

[0061] in, Indicates a point The covariance matrix of Indicates a point The covariance matrix of .

[0062] Keyframe selection defines a keyframe structure, which includes a frame of point cloud and the current robot pose. After obtaining the current frame point cloud, the current robot pose is estimated through the front-end odometry. If the current estimate exceeds a certain threshold compared to the pose transformation of the previous keyframe, the current frame is selected as the keyframe and added to the keyframe queue. Multiple keyframes constitute a local subgraph.

[0063] Frame matching matches the current frame point cloud with the local submap to reduce the robot's posture drift and error, traverses the key frame queue for matching, and finds the distance posture after the above inter-frame matching The most recent k keyframes are used as local submaps, and frame-to-frame matching also solves the pose transformation by minimizing the error of the objective function. The optimization objective function of frame-to-frame matching is as follows:

[0064]

[0065] in, Indicates a point The covariance matrix of Represents the frame s in the local submap i j The covariance matrix of .

[0066] Finally, point cloud fusion is performed, the current point cloud frame data of each machine is subscribed, point cloud downsampling is performed, and then the current point cloud and the estimated pose are accumulated to obtain the current global constructed map point cloud. Finally, the global point cloud map is published regularly for visualization and upper-level planning.

[0067] Figure 3 Experimental results of two robots collaboratively building a map in a landing experiment; Figure 4 The experimental results of three robots collaboratively building a map in a field experiment show that the multiple robots successfully collaborated to build a large-scale environment map with fewer features, and their positioning was relatively accurate, showing a good driving trajectory.

[0068] (2) Multi-robot autonomous exploration method

[0069] The overall process of the multi-robot autonomous exploration method of the present invention is as follows: Figure 5 As shown, it mainly includes the following steps:

[0070] First, the program is initialized to receive the point cloud map and positioning information obtained by multi-machine SLAM to construct the TSDF map;

[0071] Next, the cars are initialized. Each car moves forward six meters to construct the initial environment around each car's initial position, in preparation for generating an undirected graph.

[0072] Then determine whether the maximum number of iterations has been reached or whether a stop instruction has been received. If so, continue the exploration process; otherwise, jump directly to the end of the program;

[0073] Then, the global graph is expanded based on the trajectory of each vehicle and the spatial sampling around the vehicle;

[0074] Then, the free space around the car is uniformly sampled locally to construct a local graph, and the local graph and its state are output;

[0075] Determine whether the local graph state is successful. If not, proceed to global exploration. If successful, calculate the voxel gain for all local graph leaf nodes. If it is greater than the threshold, it is recorded as a local boundary node and the local graph and its state are output. Figure 6 is the computational model of node voxel gain;

[0076] Determine again whether the local graph state is successful. If not, proceed to global exploration. If successful, use the AStar algorithm to find the shortest paths between all local boundary nodes and solve the TSP problem to find the shortest access sequence, and output the local graph and its state.

[0077] Determine again whether the local graph status is successful. If not, proceed to global exploration. If successful, concatenate the paths to form the shortest path for local exploration, and add all local boundary nodes and paths as nodes and edges to the global graph, and output the final path of the local exploration.

[0078] During global exploration, the exploration gain of the global node and the voxel gain are calculated to obtain the global boundary node. At the same time, the global graphs of other robots are obtained for fusion, so that each robot no longer explores the explored area in the subsequent exploration process.

[0079] Calculate the shortest path to the nearest global boundary node using the AStar algorithm;

[0080] The number of iterations is increased by one, the path is sent to the planning controller for optimization, and the tracking path is sent to the underlying controller for tracking; the next round of iterations is continued to determine whether the maximum number of iterations has been reached.

[0081] (3) Next, we will discuss the local and global exploration separately:

[0082] 1) Multi-robot local exploration algorithm:

[0083] First, set the parameters and reset the data structure; then set the robot's expansion cube, set the robot's namespace, and pass in the robot's current position;

[0084] Next, the local graph is constructed. The space within the specified range around the sampling car is uniformly sampled to obtain a local sampling point set. The local sampling points are filtered and sorted. Occupancy points are filtered out and confidence points are stored by projecting onto the ground. Free space sampling points are recorded and added as nodes to the local graph. The local graph is expanded and collision edges are filtered out based on the car's volume expansion. The local graph is visualized and the local graph and its status are output. The local graph construction is complete.

[0085] Finally, local paths are generated, the voxel gain of each leaf node is calculated, and a path search is performed between points on the set of local boundary nodes whose gain is greater than the threshold. Here, the AStar algorithm is used to obtain the distance matrix between each local boundary node and input it into the TSP solver to solve the traveling salesman problem to obtain the shortest visit sequence. Finally, the paths are spliced ​​to obtain the final path of the local exploration.

[0086]

[0087] 2) Multi-robot global exploration algorithm:

[0088] First, after the car is initialized, a timer trigger function is constructed to periodically sample the current robot position, add it to the interval sampling points and expand it. At the same time, the collision-free area within a certain range around the current position is sampled and added to the global sampling points to expand the global graph; the local boundary nodes and the shortest paths between points are added to the global graph as global nodes and global edges; when the global graph is not empty, the voxel gain of the global node is calculated and whether it is a boundary node is determined. The exploration gain is calculated using the exponential decay exploration gain function, and the AStar algorithm is used to find the global shortest path to the global boundary node with the highest exploration gain in the global graph; if there is no boundary node in the global graph, wait for the return phase to find the path to the root node through the AStar algorithm, and the global graph exploration is completed.

[0089]

[0090]

[0091] (4) Multipath conflict detection

[0092] The main idea of ​​the multi-path conflict detection algorithm in the present invention is to split the long path into short paths and then detect whether there is overlap between the short path spaces of each pair by expanding each car; if there is overlap, the short path is sent first according to the car priority.

[0093]

[0094]

[0095] (5) Path improvement and pure tracking algorithm

[0096] After the exploration planner obtains a preliminary local path or global path, it is necessary to optimize the trajectory to be tracked and construct a pure tracking module to send speed signals to the underlying hardware.

[0097] Here we use the pure tracking module based on geometric tracking. First, we select the preview distance l of the robot's rear wheel center on the reference path. dir Reference point (g x ,g y ), assuming the robot's horizontal axis length is L, the steering radius is R, the lateral tracking error is δ, and the angle between the robot's horizontal axis and the target point is α, we can get:

[0098]

[0099]

[0100] It is concluded that the robot path tracking curvature is:

[0101]

[0102] Use the desired linear velocity and angular velocity:

[0103]

[0104]

[0105] By building a Simple_Action service for each robot to request path services from the planning manager, the pure tracking node is responsible for calculating the linear velocity and angular velocity based on the path and sending them to the underlying hardware.

[0106] Figure 7 This demonstrates the effectiveness of the multi-robot path conflict detection algorithm; Figure 8 、 Figure 9 、 Figure 10 、 Figure 11 This is a simulation process of multi-machine exploration, where the dense points are local sampling points. Figure 11 The lines are the trajectories that the robot has traveled, and finally a complete point cloud map is obtained, such as Figure 12 shown.

[0107] The above embodiments are only for illustrating the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A multi-robot autonomous exploration method, characterized in that: The steps include: Step 1: Given the initial position of each robot, initialize each robot; for each robot, move forward six meters to preliminarily construct the environment around the initial position of the robot; Step 2: determine whether the current number of iterations has reached the maximum number of iterations or whether a stop command has been received. If the maximum number of iterations has been reached or a stop command has been received, the iteration ends; otherwise, the exploration process continues; Step 3: Use the multi-machine SLAM algorithm to obtain the laser point cloud information and positioning information of each robot, and construct the TSDF map for local and global exploration; Step 4: For each robot, expand the global graph according to the robot trajectory, randomly sample global graph nodes around the robot and perform screening and expansion; Step 5: uniformly sample the local graph nodes around the robot to construct a local graph and screen obstacles, and output the local graph and its status; Step 6, determine whether the local graph state output in step 5 is successful. If so, calculate the voxel gain of all leaf nodes in the local graph, record the leaf nodes whose voxel gain is greater than the first preset threshold as the local graph boundary nodes, and output the local graph and its state; otherwise, proceed to step 9 to perform global exploration; Step 7, determine whether the local graph state output in step 6 is successful, if so, solve the traveling salesman problem for the local graph boundary nodes to obtain the shortest access sequence of the local graph boundary nodes, and output the local graph state; otherwise, go to step 9 to perform global exploration; Step 8, determine whether the local graph state output in step 7 is successful, if so, concatenate the paths according to the shortest access order and perform path optimization, use the local graph boundary nodes in the path as nodes of the global graph, use the path formed by connecting two adjacent local graph boundary nodes in the path as the edge of the global graph, and output the final path of the local exploration; otherwise, go to step 9 to perform global exploration; Step 9, calculate the exploration gain and voxel gain of all nodes in the global graph, so as to obtain the boundary nodes of the global graph, and at the same time obtain the global graph of other robots for fusion; Step 10: For each robot, use the AStar algorithm to calculate the shortest path for the robot to reach the nearest global graph boundary node; Step 11, the number of iterations is increased by one, the final path obtained by local exploration of each robot or the shortest path obtained by global exploration is optimized to obtain a tracking path, the tracking path is returned to each robot for tracking, and then step 2 is returned.

2. The multi-robot autonomous exploration method according to claim 1, characterized in that: In step 11, before optimization, a multi-path conflict detection algorithm is used to detect whether there is a conflict between the final path or the shortest path explored by each robot, that is, the priority of each robot is set in advance, and the final path or the shortest path explored by each robot is divided into short paths. Each robot is expanded to detect whether there is an overlap between the short path divided by one robot and the segment path divided by another robot. If there is an overlap, the short path corresponding to the robot with a higher level is selected according to the priority.

3. The multi-robot autonomous exploration method according to claim 1, characterized in that: In step 1, the multi-machine SLAM algorithm is used to obtain the laser point cloud information and positioning information of each robot, and the TSDF map is constructed. The specific process is as follows: (1) Perform point cloud preprocessing on each robot, delete the robot's own point cloud within a set distance around the robot, sample the point cloud at a fixed resolution and store it; (2) Performing stand-alone pose calculation for each robot, including instantaneous frame-to-frame pose estimation, key frame selection, and frame image matching; (3) Subscribe to the point cloud data currently stored by each robot, perform point cloud downsampling, and then accumulate the current point cloud of each robot, the downsampled point cloud, and the estimated pose to obtain the current global constructed map point cloud.

4. The multi-robot autonomous exploration method according to claim 3, characterized in that: The instantaneous inter-frame pose estimation is performed by comparing the point cloud of the current frame And the last frame point cloud stored Get the six-degree-of-freedom transformation T in the robot sensor coordinate system i L ; The inter-frame coordinate transformation is obtained by minimizing the error of the objective function. The objective function is as follows: in, Represents the current frame point cloud P i The 1st to nth points in Represents the point cloud P of the previous frame i-1 The 1st to nth points in e j represents the directed distance between the jth pair of points after coordinate transformation, Indicate point The covariance matrix of Indicate point The covariance matrix of , the upper right corner T represents the matrix transpose.

5. The multi-robot autonomous exploration method according to claim 3, characterized in that: The key frame selection is achieved by defining a key frame structure, which includes a frame of point cloud and the current posture of the robot. After obtaining the current frame point cloud, the current posture estimate of the robot is obtained by calculation. If the current posture estimate is transformed by more than a second preset threshold compared to the posture change of the previous key frame, the current frame is selected as the key frame and added to the key frame queue.

6. The multi-robot autonomous exploration method according to claim 5, characterized in that: The frame map matching is performed by matching the current frame point cloud with the current local submap, traversing and matching the current frame with each key frame in the key frame queue, and finding the distance pose after the traversal matching. The most recent k key frames update the local submap, and the frame map matching solves the pose transformation by minimizing the objective function error. The objective function is as follows: Among them, T i L represents the transformation of the six degrees of freedom in the robot sensor coordinate system, Represents the global coordinate transformation calculated in the previous frame in the world coordinate system, T i W Represents the global coordinate transformation calculated for the current frame in the world coordinate system, e j represents the directed distance between the jth pair of points after coordinate transformation, Indicate point The covariance matrix of Indicates the mid-frame of a local submap The covariance matrix of , the upper right corner T represents the matrix transpose.

7. A multi-robot autonomous exploration system based on the multi-robot autonomous exploration method according to any one of claims 1 to 6, characterized in that: The system comprises: a plurality of robot platforms, each robot platform comprises a robot chassis with differential wheels, a 3D mechanical laser radar, an industrial computer and an onboard router, the 3D mechanical laser radar, the industrial computer and the onboard router are all arranged on the robot chassis; the industrial computer has two network ports, the 3D mechanical laser radar transmits data with the industrial computer through one of the network ports, the onboard router transmits data with the industrial computer through the other network port, the two network ports are arranged in different network segments, the robot chassis is connected to the industrial computer through a CAN bus for data transmission; each robot platform corresponds to an ad hoc network router, each router performs ad hoc networking through Mesh; the central control terminal, the ad hoc network router, the industrial computer and the onboard router of the robot platform form a local area network; The 3D mechanical laser radar is used to obtain laser point cloud information around the robot; The industrial computer is used to deploy the multi-robot autonomous exploration method, receive control instructions from the central control terminal, and control the robot platform to obtain point cloud map information, posture information, local and global undirected graphs constructed during the robot platform exploration process, and the robot's historical trajectory; The on-board routing is used to exchange information with other robot platforms and the central control terminal; The ad hoc network router is used to expand the local area network to ensure the effectiveness of communication between each on-board router and the central control terminal; The central control terminal is used to send control instructions to the industrial control computer of each robot platform and send the short path to each robot platform after path conflict detection and segmentation. At the same time, the central control terminal receives point cloud map information, posture information, local and global undirected graphs constructed during the exploration process of each robot platform, and historical trajectory information of each robot.

Images