A distributed multi-robot autonomous collaborative exploration and mapping method

Abstract

The present invention discloses a distributed multi-robot autonomous collaborative exploration and mapping method. It includes: during the exploration process, each robot explores and obtains the global subspace state map. If each robot does not communicate with other robots, the current robot performs global exploration path planning and local exploration path planning according to the global subspace state map explored by itself, and then performs autonomous local exploration and global exploration; if each robot communicates with other robots, the current robot performs global exploration path planning and local exploration path planning combined with loop constraints according to the latest global subspace state map, and then performs collaborative local exploration and global exploration; after the exploration is completed, distributed collaborative mapping is performed to obtain positioning and mapping results. The present invention improves the autonomy and accuracy of multi-machine collaborative perception work, and has a good guiding value for the application of multi-machine collaboration in real-world scenarios.

Description

Technical Field

[0001] The present invention relates to a multi-robot collaborative exploration and mapping method in the field of robot collaborative perception, and specifically to a distributed multi-robot autonomous collaborative exploration and mapping method. Background Art

[0002] With the rapid development of robotics technology, multi-robot systems are increasingly being used in various complex environments, especially in tasks such as exploring unknown environments, positioning and mapping, and collaborative perception. Traditional single-robot systems often have problems such as low efficiency and limited data processing capabilities when facing large-scale and complex environments. For example, in scenarios such as disaster relief, large-scale environmental monitoring, or industrial automation, a single-robot system is difficult to complete a large range of tasks within a limited time and is easily disturbed by dynamic changes in the environment. Therefore, the collaborative work of multi-robot systems has become an effective way to solve these problems. Through the collaboration of multiple robots, not only can the efficiency of task execution be improved, but also the robustness and adaptability of the system can be enhanced.

[0003] In multi-robot collaborative exploration tasks, how to efficiently allocate tasks, plan paths, avoid repeated exploration, and ensure that each robot can accurately perceive and build maps are the difficulties of current research. Existing multi-robot exploration methods usually adopt a centralized control strategy, that is, a central controller uniformly plans the action paths of all robots. However, this centralized control strategy is prone to communication delays and excessive computational burdens when facing a dynamically changing environment, thus affecting the overall performance of the system. For example, in an environment with limited communication, the central controller may not be able to obtain the status information of all robots in real time, resulting in path planning lags or failures. In addition, as the number of robots and the complexity of tasks increase, the computational load of the central controller will rise sharply, which may become a performance bottleneck of the system. Therefore, distributed control strategies have gradually become a hot topic of research. By delegating decision-making power to each robot node, the communication and computational burden can be effectively reduced, the flexibility and scalability of the system can be improved, and the advantages of multi-machine collaborative systems can be fully utilized.

[0004] On the other hand, in the multi-robot collaborative positioning and mapping task, how to ensure the consistency of posture between robots and improve the mapping accuracy is also an important research direction. The traditional SLAM (Simultaneous Localization and Mapping) method mainly relies on the sensor data of a single robot. In a multi-robot system, due to the relative motion between robots and sensor noise, it is easy to cause inconsistency in pose estimation, thus affecting the final map accuracy. Therefore, how to achieve efficient loop closure detection and pose graph optimization in a multi-robot system is the key to improving the positioning and mapping accuracy. In recent years, multi-robot SLAM methods based on distributed optimization have gradually attracted attention. By introducing loop closure detection and global pose graph optimization between robots, the positioning and mapping accuracy of the system can be effectively improved. Summary of the invention

[0005] In order to overcome the deficiencies in the prior art, the purpose of the present invention is to provide a distributed multi-robot autonomous collaborative exploration and mapping method. The present invention combines the multi-robot collaborative autonomous exploration task with the mapping task. In the exploration process, the path planned by autonomous exploration is executed, and the inter-robot loop is actively introduced in a timely manner during the execution process (visiting the location visited by another robot), and then perception is performed to efficiently complete the collaborative exploration task of environmental coverage; in the collaborative mapping process, the robots perform inter-robot closed-loop detection by exchanging key frame global descriptors, and the reliability of closed-loop observation is ensured by a refined loop verification module, and reliable inter-robot closed-loop observation is added to the pose graph as an inter-robot closed-loop factor, and the global pose graph is optimized, thereby improving the accuracy of multi-robot collaborative perception. Therefore, the present invention integrates collaborative autonomous exploration and collaborative positioning and mapping tasks, so that multiple robots can autonomously and efficiently explore unknown environments and obtain accurate positioning and mapping results.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] During the multi-robot autonomous collaborative exploration process, each robot uses its own laser radar and inertial measurement unit to explore and obtain the global subspace state map. If each robot does not communicate with other robots, the current robot performs global exploration path planning and local exploration path planning based on the global subspace state map explored by itself, and then conducts autonomous local exploration and global exploration. If each robot communicates with other robots, the current robot performs global exploration path planning based on the global subspace state map explored by itself and the global subspace state map obtained through communication exchange, and uses the historical trajectories of other robots to perform local exploration path planning combined with loop constraints, and then conducts collaborative local exploration and global exploration.

[0008] After the multi-robot autonomous collaborative exploration is completed, distributed collaborative mapping is performed based on the original sensor data obtained by the multi-robots to obtain positioning and mapping results.

[0009] Each robot uses a layered environment representation method in the planning process, specifically using a low-resolution environment representation at the global level and a high-resolution environment representation at the local level. The local planning range is a set near-car area with a fixed area.

[0010] If each robot does not communicate with other robots, the current robot plans its own global exploration path based on the global subspace in the global subspace state map explored by itself using the TSP problem solver, and solves a shortest global path through all global subspaces to be explored.

[0011] If each robot communicates with other robots, the current robot updates its own global subspace state map in real time according to the global subspace state map obtained from the communication exchange, obtains the latest synchronized global subspace state map, and uses the VRP problem solver to perform global exploration path planning for different robots, obtains N different exploration paths {T1, T2, ..., TN} and assigns them to corresponding robots, so that all global subspaces to be explored in the latest global subspace state map are assigned to corresponding robots, realizing collaborative global exploration of multiple robots.

[0012] If each robot is in communication with other robots, the current robot performs nearest neighbor search and matching in the historical trajectory information of other robots acquired according to the current position, obtains the nearest matching position and uses it as a potential loop position, and uses the potential loop position as a candidate viewpoint for local planning. At the same time, the position randomly sampled by the current robot within the local planning range is also used as a candidate viewpoint for local planning, and local exploration path planning is performed based on the candidate viewpoints for local planning.

[0013] In the local exploration path planning, the utility function of the potential loop position satisfies the following formula:

[0014] A(v lc )=P(v lc )e^(-L(v0,v lc ) / t)

[0015] Among them, A(v lc ) is the potential loop position for the candidate viewpoint v lc The utility value, P(v lc ) is the potential loop position for the candidate viewpoint v lc The information gain, L(v0,v lc) is the current viewpoint v0 to the candidate viewpoint v lc , t is the duration since the last closed loop, and ^ is the power.

[0016] The distributed collaborative mapping is performed based on the original sensor data acquired by multiple robots to obtain positioning and mapping results, including:

[0017] In the collaborative mapping process of multiple robots, each robot performs lidar odometer based on the original sensor data to obtain the estimated position and radar odometer factor. Each robot extracts key frames from the input lidar point cloud data and generates the corresponding global descriptor. It also performs internal loop detection to obtain the robot's internal loop factor. During the communication process with other robots, each robot also performs inter-robot loop detection based on the global descriptor to obtain the inter-robot loop factor.

[0018] Based on the radar odometer factor, the robot's internal loop factor and the inter-robot loop factor, the estimated pose is optimized to obtain the final pose graph optimization result, thereby obtaining the positioning and mapping result.

[0019] Each robot also performs inter-robot loop detection based on the global descriptor during the communication process with other robots to obtain inter-robot loop factors, including:

[0020] Each robot communicates with other robots to exchange the global descriptor of each key frame. The global descriptor of each key frame obtained through exchange is used to perform closed-loop search in the global descriptor extracted by the local vehicle to obtain closed-loop candidates. Then, the downsampled point cloud corresponding to the closed-loop candidate is sent to other robots. The robot that obtains the downsampled point cloud through communication performs scan-to-map matching between the neighboring submap it maintains and the downsampled point cloud obtained through interaction to obtain the matching result. Finally, the RANSAC algorithm is used to verify the matching result. If the verification passes, the current matching result is precisely aligned by using the ICP algorithm to obtain the relative pose transformation and communicate the relative pose transformation back to the local robot, and the relative pose transformation is used as the loop closure factor between robots. If the verification fails, the global descriptor of the next key frame is processed until the global descriptors of all key frames are processed and the loop closure factors between all robots are obtained.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. The present invention realizes the collaborative and autonomous environmental perception of multiple robots without the need for human control and in an unknown environment, and can autonomously explore and complete collaborative positioning and mapping.

[0023] 2. The present invention uses the data obtained from autonomous exploration to perform collaborative mapping, organically combines the two tasks, and actively introduces a loop mechanism between robots, so that the robot introduces loop constraints based on visiting locations visited by other robots, thereby improving the accuracy of robot positioning and mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solution of the embodiment of the present invention and better reflect the innovation and practicality of the invention as well as the basic technical principles, the present invention is further described in detail below with reference to the accompanying drawings.

[0025] Figure 1 The figure is a flow chart of the overall method of the present invention.

[0026] Figure 2 This is a flowchart of the local exploration path planning combined with loop constraints proposed by the present invention.

[0027] Figure 3 Schematic diagram of a multi-machine autonomous collaborative exploration process according to an embodiment of the present invention.

[0028] Figure 4 Schematic diagram of a multi-machine collaborative mapping process according to an embodiment of the present invention.

[0029] Figure 5 Schematic diagram of multi-machine collaborative mapping results according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0031] like Figure 1 The specific implementation process is as follows:

[0032] During the multi-robot autonomous collaborative exploration process, each robot uses its own laser radar and inertial measurement unit (IMU) to explore and obtain the global subspace state map. Whenever two or more robots are within the communication range, each robot can exchange their own global map information with other robots within the communication range, and update their own global subspace information to obtain a more complete and real-time global map, that is, the global subspace state map. If each robot does not communicate with other robots, the current robot will perform global exploration path planning and local exploration path planning based on the global subspace state map explored by itself, and then conduct autonomous local exploration and global exploration; if each robot communicates with other robots, the current robot will perform global exploration path planning based on the global subspace state map explored by itself and the global subspace state map obtained through communication exchange, and use the historical trajectories of other robots to perform local exploration path planning combined with loop constraints, and then conduct collaborative local exploration and global exploration;

[0033] The robot performs exploration in an unknown environment and performs exploration planning tasks based on sensor information, that is, planning the self-vehicle exploration path. Each robot uses a hierarchical environment representation method in the planning process, specifically using a low-resolution (sparse) environment representation at the global level and a high-resolution (dense) environment representation at the local level. The main computing resources are concentrated in the space close to the robot, rather than wasted in places with greater uncertainty at a distance. At the same time, because of the existence of global path planning, the robot's driving path will be guided by global information and more purposeful. The local planning range is a set near-car area with a fixed area. That is, within the local planning area (near-car area), the robot will drive along a "fine" path given by a large number of calculations, while outside this area, the robot will be guided by a "rough" path given by a small number of calculations in the general direction of its driving. During the exploration process, the environment is divided into three-dimensional cubic spaces of fixed size, called global subspaces. The global subspace can be divided into three states: unexplored, explored, and to be explored according to the sensor's perception information, and the state is continuously updated with the input of sensor data. The state of the global subspace is mainly determined by the state of the generalized surface located therein, which not only considers the boundary between free space and non-free space, but also the boundary of the object surface. The surface sensed by the sensor within a certain distance and angle range is set to be in a covered state, and the surface that does not meet the distance and angle conditions is sensed but not covered. If there are only surfaces in the covered state in a global subspace, it is in a state of exploration; if there are both surfaces in the covered state and surfaces in the sensed state in a global subspace, it is in a state to be explored; if there are no surfaces in the covered state or sensed state in a global subspace, it is in an unexplored state. The problem to be solved by local planning is to maximize the surface coverage problem, and to find the shortest path in the local area that allows the sensor to "see everything and look forward to it", so that the robot can completely cover the surface in the environment along this path and complete comprehensive and complete environmental perception. Figure 3 (a) and Figure 3 The small solid green blocks in (b) are low-resolution global subspaces to be explored (i.e., places in the global environment that are not fully covered, and the robot will need to go to these locations in the future to cover the environment completely). The large boxes are local areas currently being explored, which contain multiple global subspaces, in which refined local planning is performed so that the sensor can fully cover the local area.

[0034] If each robot does not communicate with other robots, the current robot plans its own global exploration path based on the global subspace in the global subspace state map explored by itself using the TSP (Traveling Salesman Problem) problem solver. The solution obtains the shortest global path through all the global subspaces to be explored, so as to ensure that the robot can efficiently explore the environment and complete the perception of the environment.

[0035] If each robot communicates with other robots, the current robot updates its own global subspace state map in real time based on the global subspace state map obtained through communication exchange, obtains the latest synchronized global subspace state map, and uses the VRP (Vehicle Routing Problem) problem solver to perform global exploration path planning for different robots, obtains N different exploration paths {T1, T2,…, TN} and assigns them to the corresponding robots, and minimizes the maximum value of the length l of the N different exploration paths. By default, all robots start exploring at the same time, and the maximum value of the length l of the N different exploration paths can represent the end time of the exploration to a certain extent, so that all the global subspaces to be explored in the latest global subspace state map are assigned to the corresponding robots, realizing the collaborative global exploration of multiple robots. The planning objective function satisfies the following formula:

[0036] argmin T1,…,Tn max l(Ti), i∈{1,2,...,N}

[0037] Among them, T1 and TN are the planned global exploration paths corresponding to the 1st and Nth robots respectively, l(Ti) is the length of the global exploration path Ti of the ith robot, and the maximum value among them is obtained according to the lengths of the N paths. Solving the VRP problem is to make this maximum value as small as possible and to minimize the total exploration time, so as to efficiently complete the collaborative exploration task.

[0038] like Figure 2As shown in the figure, if each robot communicates with other robots, the current robot (i.e., the local robot) performs nearest neighbor search and matching in the historical trajectory information of other robots according to the current position information, obtains the nearest matching position and uses it as a potential loop position. The potential loop position is used as a candidate viewpoint for local planning. At the same time, the position randomly sampled by the current robot within the local planning range (near the vehicle area) is also used as a candidate viewpoint for local planning. According to the proposed utility function, a suitable viewpoint is selected from the candidate viewpoints based on local planning as the local exploration path under the comprehensive consideration of information gain and path cost. In the local planning process, the selection of the candidate viewpoints is mainly based on the utility function. The utility function of the non-potential loop position satisfies the following formula:

[0039] A(v)=P(v) e^(-L(v0,v))

[0040] Among them, A(v) is the utility value of the non-potential loop position as the candidate viewpoint v, P(v) is the information gain of the non-potential loop position as the candidate viewpoint v, that is, the range of the unknown area that the candidate viewpoint v can cover, and L(v0,v) is the path cost from the current viewpoint v0 to the candidate viewpoint v. The process of selecting the viewpoint is to select the viewpoint position that maximizes the utility function from the randomly sampled viewpoint candidates according to the utility function that maximizes the information gain and path cost.

[0041] In local exploration path planning, the utility function of the potential loop position satisfies the following formula:

[0042] A(v lc )=P(v lc )e^(-L(v0,v lc ) / t)

[0043] Among them, A(v lc ) is the potential loop position for the candidate viewpoint v lc The utility value, P(v lc ) is the potential loop position for the candidate viewpoint v lc The information gain, L(v0,v lc ) is the current viewpoint v0 to the candidate viewpoint v lcThe path cost is t, t is the duration since the last loop closure. If it is the first loop closure, it is the duration from the beginning of exploration to the current position, and ^ is a power. This utility function aims to select the candidate viewpoint that maximizes the coverage of the unknown area and minimizes the path cost as the target viewpoint, and plan the robot to go to this position for perception. The effect of the duration t since the last loop closure is added to the path cost term. As the duration t increases, the utility function value of the potential loop viewpoint continues to increase. When the loop is not closed for a long time, the utility value of the potential loop position viewpoint is larger, which encourages the planning method to plan a path through the potential loop viewpoint v lc , so as to actively introduce the loop between robots, such as Figure 2 shown. Figure 3 The visualization shows the process of collaborative exploration of the two robots. It can be seen that the two robots come to different areas in the environment for exploration under the planning of the method of the present invention.

[0044] For a subspace within the local planning range, the exploration of the subspace is completed only when all surfaces in the space are covered. The surface here is defined as a generalized surface, including both the boundary between free space and non-free space and the surface of objects. When the system is planning locally, it will look for a path that maximizes the coverage of the surface of the object by the sensor field of view. In the local area, by cyclically sampling viewpoints, the shortest path that can maximize the coverage of the surface is found.

[0045] After the multi-robot autonomous collaborative exploration is completed, distributed collaborative mapping is performed based on the original sensor data (including lidar point cloud data and IMU data) obtained by the multi-robots to obtain positioning and mapping results.

[0046] Distributed collaborative mapping is performed based on the original sensor data (including lidar point cloud data and IMU data) acquired by multiple robots to obtain positioning and mapping results, including:

[0047] In the collaborative mapping process of multiple robots, each robot performs lidar odometer according to the original sensor data to obtain the estimated pose and radar odometer factor. Each robot extracts key frames from the input lidar point cloud data and generates the corresponding global descriptor. In this embodiment, the LiDAR-Iris global descriptor with rotation invariance is used, which makes full use of most of the information of the point cloud while lightweight representation of point cloud features, and also avoids brute force search to save computing resources. In order to improve the accuracy of the single robot front end, loop detection is also performed inside the robot to obtain the internal loop factor of the robot; each robot also performs inter-robot loop detection based on the global descriptor during communication with other robots to obtain the inter-robot loop factor; these factors obtained in the collaborative mapping process are all used in pose graph optimization. The lidar odometer front end used in this embodiment is LIO-SAM, which combines the high-precision distance measurement of lidar and the posture estimation of IMU, and can achieve high-precision positioning and map construction in complex environments.

[0048] The pose graph optimization process optimizes the estimated pose based on the radar odometer factor, robot internal loop factor and inter-robot loop factor obtained by the above operations to obtain the final pose graph optimization result, thereby obtaining the multi-robot collaborative positioning and mapping result and completing efficient and accurate perception of the environment.

[0049] During the communication process between each robot and other robots, loop closure detection is performed between robots based on the global descriptor to obtain the loop closure factor between robots, including:

[0050] Each robot communicates with other robots to exchange the global descriptor of each key frame. The global descriptor of each key frame obtained by the exchange is used to perform closed-loop search in the global descriptor extracted by the local self-vehicle to obtain closed-loop candidates. Then, the downsampled point cloud corresponding to the closed-loop candidate is sent to other robots. Then, the neighboring submap of the robot that obtains the downsampled point cloud through communication and the downsampled point cloud obtained through interaction are matched by the scan-to-map matching method to obtain the matching result. Finally, the RANSAC algorithm is used to verify the matching result. If the verification is passed, the current matching result is precisely aligned by using the ICP algorithm to obtain the relative pose transformation and the relative pose transformation is communicated back to the local robot, and the relative pose transformation is added to the factor graph as the inter-robot loop factor; if the verification fails, the global descriptor of the next key frame is processed until the global descriptors of all key frames are processed and all inter-robot loop factors are obtained. Among them, the matching result with a sufficient number of internal points in the RANSAC algorithm is considered to be verified, that is, the closed-loop candidate is considered to be reliable, and the closed-loop observation obtained by ICP is reliable, which can be used for subsequent pose graph optimization.

[0051] The general likelihood formula including the radar odometry factor, the robot internal loop closure factor and the inter-robot loop closure factor is shown in the following formula:

[0052] Φ(x) = ψ(z αiβj | x)

[0053] x = [x α , x β , x γ ,…]

[0054] Among them, Φ(x) is a general likelihood function representation (here taking the inter-robot loop factor as an example), ψ( ) represents the inter-robot closed-loop observation z for a given position x. αiβj The likelihood probability, x is a series of robot trajectory postures (i.e., postures at each moment), x α , x β , x γ are the robot’s positions at time α, β, and γ respectively, αiβj It is the posture transformation observation of robot α at time i and robot β at time j, that is, the reliable closed-loop observation result obtained by the above-mentioned inter-robot closed-loop detection module.

[0055] The pose graph optimization process estimates the robot's trajectory by solving the maximum likelihood problem of the following formula based on laser odometry observations and closed-loop observations (inside the robot and between robots):

[0056] x'=argmax∏Φ(x)

[0057] Among them, x' is the trajectory pose result obtained after pose graph optimization. The goal of pose graph optimization is to find a set of poses x that maximizes the joint likelihood function (that is, the likelihood functions corresponding to different factors are optimized together).

[0058] During the pose graph optimization process, the necessary rotation and posture estimates are transmitted to the specified robot according to the robot's optimization order. In this way, it is possible to avoid repeated calculations and achieve global consistency on the final optimized trajectory estimate while exchanging a small amount of information.

[0059] Figure 4 The process of collaborative mapping by two robots is shown. The red and green tracks in the figure represent the tracks that the two robots have traveled, and the red and green point clouds represent the point cloud models of the maps they have built. From the results in the figure, we can see that the two robots explored different areas of the environment and completed autonomous collaborative exploration of the environment. Figure 5It is a complete point cloud model obtained by the collaborative mapping of the two, which is distinguished by different colors. The point cloud map models obtained by the two have no obvious drift, and can complement and fully characterize the unknown environment to be explored. The overlapping point cloud maps of the two can basically correspond to each other without drift or blank omissions, which qualitatively shows that the environmental model obtained by the collaborative mapping of the two is complete and accurate.

[0060] Table 1 lists the comparison of the positioning results of the two robots' individual mapping and collaborative mapping under the simulation environment model.

[0061] Table 1 is a comparison of the positioning errors (m) of the present invention under the simulation environment model

[0062]

[0063] Depend on Figure 4 As can be seen from Table 1, the robot can complete the exploration and mapping tasks efficiently with or without a closed loop. However, when the closed loop constraint is introduced, the positioning error of the robot is improved. The positioning error of robot A is improved more significantly.

[0064] Finally, it should be noted that the above embodiments and explanations are only used to illustrate the technical solution of the present invention rather than to limit it. Those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope disclosed in the technical solution of the present invention, which should be included in the scope of protection of the claims of the present invention.

Claims

1. A distributed multi-robot autonomous collaborative exploration and mapping method, characterized in that: The steps include: During the multi-robot autonomous collaborative exploration process, each robot uses its own laser radar and inertial measurement unit to explore and obtain the global subspace state map. If each robot does not communicate with other robots, the current robot performs global exploration path planning and local exploration path planning based on the global subspace state map explored by itself, and then conducts autonomous local exploration and global exploration. If each robot communicates with other robots, the current robot performs global exploration path planning based on the global subspace state map explored by itself and the global subspace state map obtained through communication exchange, and uses the historical trajectories of other robots to perform local exploration path planning combined with loop constraints, and then conducts collaborative local exploration and global exploration. After the multi-robot autonomous collaborative exploration is completed, distributed collaborative mapping is performed based on the original sensor data obtained by the multi-robots to obtain positioning and mapping results; If each robot is in communication with other robots, the current robot performs nearest neighbor search and matching in the historical trajectory information of other robots acquired according to the current position, obtains the nearest matching position and uses it as a potential loop position, uses the potential loop position as a candidate viewpoint for local planning, and at the same time, the position randomly sampled by the current robot within the local planning range is also used as a candidate viewpoint for local planning, and local exploration path planning is performed based on the candidate viewpoints for local planning; In the local exploration path planning, the utility function of the potential loop position satisfies the following formula: A(v lc )=P(v lc )e^(-L(v0,v lc ) / t) Among them, A(v lc ) is the potential loop position for the candidate viewpoint v lc The utility value, P(v lc ) is the potential loop position for the candidate viewpoint v lc The information gain, L(v0,v lc ) is the current viewpoint v0 to the candidate viewpoint v lc , t is the duration since the last closed loop, and ^ is the power.

2. A distributed multi-robot autonomous collaborative exploration and mapping method according to claim 1, characterized in that: Each robot uses a hierarchical environment representation method during the planning process, specifically using a low-resolution environment representation at the global level and a high-resolution environment representation at the local level.

3. A distributed multi-robot autonomous collaborative exploration and mapping method according to claim 1, characterized in that: If each robot does not communicate with other robots, the current robot plans its own global exploration path based on the global subspace in the global subspace state map explored by itself using the TSP problem solver, and solves a shortest global path through all global subspaces to be explored.

4. A distributed multi-robot autonomous collaborative exploration and mapping method according to claim 1, characterized in that: If each robot communicates with other robots, the current robot updates its own global subspace state map in real time according to the global subspace state map obtained from the communication exchange, obtains the latest synchronized global subspace state map, and uses the VRP problem solver to perform global exploration path planning for different robots, obtains N different exploration paths {T1, T2, ..., TN} and assigns them to corresponding robots, so that all global subspaces to be explored in the latest global subspace state map are assigned to corresponding robots, realizing collaborative global exploration of multiple robots.

5. The distributed multi-robot autonomous collaborative exploration and mapping method according to claim 1, characterized in that: The distributed collaborative mapping is performed based on the original sensor data acquired by multiple robots to obtain positioning and mapping results, including: In the collaborative mapping process of multiple robots, each robot performs lidar odometer based on the original sensor data to obtain the estimated position and radar odometer factor. Each robot extracts key frames from the input lidar point cloud data and generates the corresponding global descriptor. It also performs internal loop detection to obtain the robot's internal loop factor. During the communication process with other robots, each robot also performs inter-robot loop detection based on the global descriptor to obtain the inter-robot loop factor. Based on the radar odometer factor, the robot's internal loop factor and the inter-robot loop factor, the estimated pose is optimized to obtain the final pose graph optimization result, thereby obtaining the positioning and mapping result.

6. A distributed multi-robot autonomous collaborative exploration and mapping method according to claim 5, characterized in that: Each robot also performs inter-robot loop detection based on the global descriptor during the communication process with other robots to obtain inter-robot loop factors, including: Each robot communicates with other robots to exchange the global descriptor of each key frame, and uses the global descriptor of each key frame obtained through exchange to perform closed-loop search in the global descriptor extracted locally to obtain closed-loop candidates, and then sends the downsampled point cloud corresponding to the closed-loop candidate to other robots. The robot that obtains the downsampled point cloud through communication performs scan-to-map matching between the neighboring submap it maintains and the downsampled point cloud obtained through interaction to obtain the matching result, and finally uses the RANSAC algorithm to verify the matching result. If the verification passes, the current matching result is precisely aligned by using the ICP algorithm to obtain the relative pose transformation and communicate the relative pose transformation back to the local robot, and the relative pose transformation is used as the loop closure factor between robots. If the verification fails, the global descriptor of the next key frame is processed until the global descriptors of all key frames are processed and the loop closure factors between all robots are obtained.

Images