A robot autonomous exploration mapping system and method

Abstract

The application discloses a kind of robot autonomous exploration mapping system and method, system includes perception module, decision module, planning module and control module, the output of perception module is connected with the input of decision module, and perception module transmits front point candidate set to decision module, the output of decision module is connected with the input of planning module, and decision module transmits optimal target point to planning module, the output of planning module is connected with the input of control module, and planning module transmits motion instruction to control module, and the output of control module is connected with the input of perception module, and control module state information is fed back to perception module, and forms closed loop control system. It can be adapted to diversified industrial scene, with the advantages of efficient completion of autonomous exploration, balancing decision quality and real-time, self-adapting different structure environment, improving decision stability and the like.

Description

Technical Field

[0001] This invention relates to the field of robot navigation and mapping technology, and in particular to a robot autonomous exploration mapping system and method. Background Technology

[0002] A robot's ability to perceive its own structure (including its mechanical structure and external physical structure) is the cornerstone of its intelligent behaviors, such as autonomous navigation, precise operation, and safe collaboration. In industrial scenarios such as nuclear power plant inspection, warehousing and logistics, and construction, mobile robots often need to continuously build maps in unknown or semi-unknown environments. Autonomous exploration, as a core component of these robots' work, is crucial for accurately selecting the next target point online, quickly reducing environmental uncertainty, and ensuring the orderly progress of subsequent inspections and operations. However, existing autonomous exploration technologies still have many key bottlenecks, making it difficult to meet the actual needs of industrial scenarios. Therefore, intelligent autonomous exploration methods are gradually becoming a research hotspot.

[0003] CN116896723A discloses an autonomous exploration and obstacle avoidance method based on adaptive frontier exploration target point selection. By adaptively adjusting thresholds and combining obstacle avoidance priority point selection, it achieves accurate point selection and obstacle avoidance, adapting to simple industrial structured environments. However, it does not address the core shortcomings of traditional methods, failing to explicitly express exploration value, exhibiting path oscillations in typical scenarios, and not considering computational limitations, making it difficult to adapt to embedded platforms and meet real-time requirements. CN117405688A discloses a robot navigation and interactive control method and system based on multi-map fusion. It fuses point cloud files based on scene topology relationships and combines topological features for point selection and navigation, solving problems of insufficient structure utilization and low cross-scene exploration efficiency. However, it relies on strong prior knowledge and heavy semantic modeling, resulting in high computational complexity, poor real-time performance, difficulty adapting to industrial mobile robots, and a tendency to introduce biases in weakly structured environments, with insufficient optimization for autonomous exploration point selection. CN118181571A discloses a mobile robot autonomous exploration method based on information gain optimization. It combines information gain and path cost evaluation, introduces a greedy strategy for point selection to improve globality and rationality, and solves problems inherent in traditional information theory methods. It is suitable for small- to medium-scale industrial scenarios, but suffers from high computational overhead, poor stability, heavy planning pressure on large scales, and a tendency to trigger replanning blocking. It is not optimized for weakly structured environments and is difficult to adapt to complex scenarios. CN119000241A discloses a robot autonomous exploration and navigation method based on deep learning. It uses a deep learning model to predict information gain and introduces transfer learning to reduce dependence on training data, improving the accuracy of point selection in specific structured industrial environments. However, it is highly dependent on training data, has limited environment generalization ability, and suffers from model inference latency, making it difficult to meet real-time requirements and adapt to embedded platforms.

[0004] Unlike previous research on autonomous exploration methods for mobile robots, previous patents have not specifically addressed the issues of balancing real-time performance, stability, and environmental adaptability in unknown or semi-unknown industrial environments. Existing related patents all have their limitations and cannot effectively meet the complex and diverse exploration needs in industrial scenarios.

[0005] Based on the technical deficiencies of the existing patents, the existing autonomous exploration technology for mobile robots has four major common problems in industrial applications, which seriously restrict its engineering applicability: insufficient structural utilization, contradiction between real-time computing and decision quality, poor decision stability, and insufficient adaptability to weak structure scenarios. Summary of the Invention

[0006] Purpose of the invention: To address the above-mentioned problems, the purpose of this invention is to provide a robot autonomous exploration mapping system and method that can be adapted to diverse industrial scenarios and has advantages such as efficient autonomous exploration, balancing decision quality and real-time performance, adapting to different structural environments, and improving decision stability.

[0007] Technical Solution: A robot autonomous exploration mapping system includes a perception module, a decision-making module, a planning module, and a control module. The output of the perception module is connected to the input of the decision-making module. The perception module transmits the candidate set of leading edge points to the decision-making module. The output of the decision-making module is connected to the input of the planning module. The decision-making module transmits the optimal target point to the planning module. The output of the planning module is connected to the input of the control module. The planning module transmits motion commands to the control module. The output of the control module is connected to the input of the perception module. The control module feeds back state information to the perception module, forming a closed-loop control system.

[0008] Furthermore, the perception module includes a LiDAR sensor, a grid map building component, and a leading edge detection component. The LiDAR sensor is used to collect environmental distance information. The grid map building component is connected to the LiDAR sensor signal, converts the distance information into an occupied grid map, and calculates the map entropy to measure the degree of environmental unknown. The leading edge detection component is connected to the grid map building component signal. The leading edge detection component is used to detect the boundary grid between unknown areas and known free areas, generates a candidate point set through clustering, and transmits the information to the decision module.

[0009] Furthermore, the decision-making module includes a structure perception calculation component, an information gain calculation component, a navigation cost calculation component, and a Top-K hierarchical screening component. The structure perception calculation component, information gain calculation component, and navigation cost calculation component calculate the structure perception score, information gain, and navigation cost of each candidate point in parallel, and transmit the calculation results to the Top-K hierarchical screening component. The Top-K hierarchical screening component performs a two-stage screening process of coarse screening and fine screening based on the received three scores, and outputs the optimal target point to the planning module.

[0010] Ideally, the planning module includes a global path planning component and a local trajectory tracking component for signal connections. The global path planning component performs path planning on candidate points to obtain the true cost; the local trajectory tracking component generates a smooth local trajectory based on the global path.

[0011] Ideally, the control module includes a motion control component and a state feedback component. The motion control component is signal-connected to the planning module, and the state feedback component transmits the robot's pose information to the perception module.

[0012] A method for performing autonomous exploration mapping using the aforementioned robot autonomous exploration mapping system includes the following steps:

[0013] Step S1: Environmental characterization and map construction;

[0014] The sensor module collects environmental information, generates an occupied grid map, and calculates the map entropy H(M) to measure the degree of environmental uncertainty.

[0015] Step S2: Extraction of frontier candidate points;

[0016] The perception module detects unknown-free region boundary grids on the grid map, clusters the leading edge points and calculates the geometric center of each region to generate a candidate point set, while removing leading edge regions with fewer than 3 leading edge points to suppress noise; Step S3: Calculation of multi-objective utility index;

[0017] The decision module obtains the distance field at each candidate point and calculates the structure perception score. ;

[0018] The decision module constructs a set of UFs on the occupied grid map, performs omnidirectional visibility analysis centered on candidate points and calculates information gain, and then normalizes the results to obtain the frontier scale weights. ;

[0019] The decision module calculates the path cost. If global planning times out or fails, an approximation will be made using Euclidean distance and the proportion of local obstacles.

[0020] Step S4: Construction of a multi-objective utility model;

[0021] According to the obtained , , Calculate the overall utility score for each candidate point. ;

[0022] Step S5: Top-K two-stage stratified screening;

[0023] After performing security checks and filtering on the candidate point set, the decision-making module calculates a coarse-screen utility score using Euclidean distance as an approximation cost, and selects the top K candidate points in descending order. The planning module then obtains the actual path costs for each of the K candidate points, calculates a fine-screen utility score, and selects the one with the highest utility. As the optimal target point;

[0024] Step S6: Target selection and motion execution;

[0025] The optimal target point The data is transmitted to the planning module to generate a global path and a smooth local trajectory. The control module then controls the robot to perform motion according to the differential drive model and feeds back the pose information to the perception module in real time.

[0026] Step S7: Candidate point set empty determination and loop control;

[0027] Perform front-edge detection on the grid map. If the candidate point set is not empty, reset the consecutive no-front-edge counter and return to step S1 to continue exploration. If the candidate point set is empty, check whether the robot's current position is near the topological critical region. If so, actively trigger map update and recheck after a 0.5-second delay. Otherwise, increment the consecutive no-front-edge counter. When the consecutive empty detection reaches 2 times, determine that the exploration is complete and end the process.

[0028] Furthermore, in step S1, the formula for calculating H(M) is:

[0029] ;

[0030] in, For grid cells The probability of being occupied. This represents the summation over all raster cells, specifically map entropy. The larger the value, the higher the degree of uncertainty in the environment.

[0031] Furthermore, in step S3, a range field is constructed based on the Euclidean distance transformation (EDT). The mathematical model is:

[0032] ;

[0033] in For the set of occupied / obstacle grids, This indicates that the minimum value is taken; the structure-aware item first filters candidate points through the bottleneck criterion, and only uses continuous scoring of channel width when conditions such as connectivity and geometric centrality are met; otherwise, it is processed according to the baseline score.

[0034] Channel width continuous scoring function for:

[0035] ;

[0036] in The channel width at the candidate point. and These represent the lower and upper limits of the ideal channel width range. For narrow channel threshold, For wide channel threshold, To score for narrow channels, Scoring is given for open or weakly structured areas;

[0037] Bottleneck criteria include: the point is not close to the map edge, the number of obstacles within the window is within a reasonable range, the number of local obstacles is moderate, and the estimated channel width falls within the configured bottleneck width range.

[0038] The UF set is a set of boundary grids that simultaneously satisfy high entropy, high gradient, and are located within a known region, denoted as:

[0039] ;

[0040] in For grid The entropy value, Entropy threshold Let be the magnitude of the entropy gradient. For gradient threshold, A function to determine whether a grid belongs to a known region;

[0041] The known region refers to a grid cell with a determined probability value, including occupied grid cells and free grid cells; uncertainty frontier They are typically located at the boundary between known and unknown areas, representing potential observation locations where sensors may acquire new information;

[0042] With candidate points Perform omnidirectional visibility analysis around the center and accumulate visibility. raster entropy, information gain The representation is:

[0043] ;

[0044] in To start from candidate points Visible The grid set; the final utility, after normalization and the introduction of frontier scale weights, is expressed as:

[0045] ;

[0046] in The maximum information gain of all candidate points. The size of the frontier region corresponding to the candidate point. For frontier size weighting coefficients;

[0047] The path cost is a weighted combination of path length, path sparsity, and turning complexity, expressed as:

[0048] ;

[0049] in For path length, For path sparsity, For turning complexity, , , These are the corresponding weights; the path sparsity is calculated by statistically analyzing the proportion of obstacle grids within the neighborhood window of each sampling point on the planned path, and is used to characterize the congestion level of the surrounding environment of the path; the turning complexity is defined as the sum of the absolute values ​​of the turning angles between adjacent road segments on the planned path.

[0050] Ideally, in step S4, the multi-objective utility evaluation includes three parallel branches: structural feature quantification, information gain assessment, and navigation risk assessment. Finally, a candidate point score is calculated through comprehensive utility calculation. The structural perception term, information gain term, and navigation cost term are weighted and fused to construct a comprehensive utility scoring function, expressed as:

[0051] ;

[0052] in For the weights of the structure-aware terms, For the information gain term weight, This refers to the weight of the navigation cost term.

[0053] Ideally, in step S5, the two-stage stratified screening is divided into coarse screening and fine screening. The coarse screening first examines the candidate point set. Perform security checks and filtering to remove unreachable candidate points or those located near obstacles, resulting in a safe candidate set. Then, the path cost is approximated using Euclidean distance to quickly compress the candidate set, and the Euclidean distance is normalized using an absolute benchmark; the utility score of the coarse screening is expressed as:

[0054] ;

[0055] in The cost of the normalized Euclidean distance approximation is expressed as:

[0056] ;

[0057] in This is the robot's current position. Candidate points The Euclidean distance to the robot's current position. As the baseline value for cost; for Candidate points Sort in descending order and extract the first few truncations. The candidate points are selected for fine screening. The computational complexity of the coarse screening is O(n log n). ;

[0058] Fine screening is only for Perform real-world planning assessments on each candidate site. Perform global path planning on each candidate point, obtain the actual path cost, and map it to... to The interval, the true path cost is represented as:

[0059] ;

[0060] in For the cost of planning the path, The baseline cost is used; the final utility score for the refined screening is expressed as:

[0061] ;

[0062] Among them, the cost item Adopting real planning costs As shown in formula (12):

[0063] Select the candidate point with the highest utility score as the optimal target point, denoted as:

[0064] ;

[0065] The computational complexity of fine screening is ,in The computational complexity of a single path planning operation is ; the total computational complexity is . This is lower than performing exact global planning on all candidate points. Complexity; Number of Top-K selections for to Any integer within the range.

[0066] Beneficial effects: Compared with the prior art, the advantages of the present invention are:

[0067] 1. This invention employs a structure perception term based on Euclidean distance transformation. By constructing a distance field and designing a continuous scoring function for channel width, it can effectively identify topologically critical locations such as narrow connected structures and structural intersections in the environment. This enables the robot to prioritize exploring key connected locations that may connect to unknown areas, reducing repeated visits and path oscillations.

[0068] 2. This invention adopts a Top-K hierarchical screening strategy. The coarse screening uses Euclidean distance to approximate the path cost and quickly compresses the candidate set. The fine screening only performs real planning evaluation on a small number of candidate points, thereby reducing the computational complexity from O(N×P) to O(N+K×P), which significantly reduces computational overhead and decision delay.

[0069] 3. This invention employs an information gain calculation method based on uncertainty frontier (UF). By constructing a set of boundary grids on the occupied grid map that simultaneously satisfy high entropy, high gradient, and are located within a known region, and performing omnidirectional visibility analysis centered on the candidate point to accumulate the entropy value of the visible UF grids, this invention more accurately measures the expected information gain of the robot after reaching the candidate point and improves information acquisition efficiency.

[0070] 4. This invention employs a dynamic weight adjustment mechanism. When the number of available candidate points is lower than a threshold, the information gain weight is automatically reduced and the cost term weight is increased. This makes the strategy more biased towards cost-sensitive neighboring regions in the later stages of exploration, avoiding unnecessary long-distance exploration when information gains diminish, and improving adaptability.

[0071] 5. This invention employs a structural term degradation mechanism. When the environmental structural features are not obvious, the structural terms naturally degrade to the baseline level, thereby avoiding the introduction of inappropriate structural biases in weakly structured scenarios and improving scenario adaptability.

[0072] 6. This invention employs a multi-objective optimization method that unifies the structure-aware term, information gain, and cost term into the same scoring function, thereby achieving a comprehensive balance between exploration efficiency, computational real-time performance, and decision stability. Attached Figure Description

[0073] Figure 1 This is a system composition diagram of the present invention;

[0074] Figure 2 This is a schematic diagram of the structural field of the present invention;

[0075] Figure 3 This is a flowchart of the robot autonomous exploration mapping method of the present invention;

[0076] Figure 4 This is a flowchart of the SUR multi-objective utility evaluation process of the present invention;

[0077] Figure 5 A flowchart for hierarchical screening of candidate points based on Top-K;

[0078] Figure 6 This is a rendering of the robot's autonomous exploration and mapping capabilities according to the present invention. Detailed Implementation

[0079] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0080] A robot autonomous exploration and mapping system, such as Figure 1 As shown, the system includes a perception module 1, a decision-making module 2, a planning module 3, and a control module 4. The perception module 1 includes a LiDAR sensor 11, a grid map construction component 12, and a leading edge detection component 13. The LiDAR sensor 11 is used to collect environmental distance information; the grid map construction component 12 is used to convert the distance information into an occupied grid map and calculate map entropy to measure the degree of environmental uncertainty; the leading edge detection component 13 is used to detect the boundary grids between unknown areas and known free areas, and to generate a candidate point set through clustering.

[0081] Decision module 2 includes a structure-awareness calculation component 21, an information gain calculation component 22, a navigation cost calculation component 23, and a Top-K hierarchical screening component 24. The structure-awareness calculation component 21 identifies topologically critical locations such as narrow connected structures and structural intersections based on Euclidean distance transformation; the information gain calculation component 22 accumulates the entropy value of the visible uncertainty frontier through ray tracing; the navigation cost calculation component 23 evaluates the weighted combination of path length, sparsity, and turning complexity; and the Top-K hierarchical screening component 24 adopts a two-stage strategy of coarse screening and fine screening to reduce computational complexity.

[0082] The planning module 3 includes a global path planning component 31 and a local trajectory tracking component 32. The global path planning component 31 performs path planning on candidate points to obtain the true cost; the local trajectory tracking component 32 generates a smooth local trajectory based on the global path.

[0083] The control module 4 includes a motion control component 41 and a state feedback component 42. The motion control component 41 controls the robot's movement according to the plan output; the state feedback component 42 feeds back information such as position, speed, and heading to the decision module.

[0084] The output of the sensing module 1 is connected to the input of the decision module 2, transmitting the candidate set of leading edge points to the decision module 2; the output of the decision module 2 is connected to the input of the planning module 3, transmitting the optimal target point to the planning module 3; the output of the planning module 3 is connected to the input of the control module 4, transmitting motion commands to the control module 4; the output of the control module 4 is connected to the input of the sensing module 1, feeding back state information to the sensing module 1, forming a closed-loop control system.

[0085] As an example, the robot autonomous exploration mapping process of the present invention includes core technologies such as environmental characterization and map construction, frontier candidate point extraction, structural feature quantification, information benefit and risk assessment, multi-objective utility model construction, Top-K two-stage hierarchical screening, target selection and motion execution, candidate point set emptying judgment and loop control, etc. The specific principles are explained as follows.

[0086] The environmental characterization and map building described herein are based on the ROS 2 (Humble) and Nav2 frameworks, utilizing a 360° two-dimensional LiDAR mounted on a differential robot to construct an occupied grid map in real time. The LiDAR sensor 11 is used to collect distance information in the environment, providing basic data for subsequent map building and obstacle detection. The robot's moving speed and angular velocity are determined by the navigation controller according to dynamic obstacle avoidance and path tracking requirements, and the sensor's scanning range covers the entire surrounding environment. The grid map building component 12 converts the distance information collected by the LiDAR sensor into an occupied grid map, using this map to characterize the environment. Each grid cell in the map represents the probability that the location is occupied by an obstacle. Grid states are classified into three categories based on probabilistic Bayesian inference results: occupied grids, free grids, and unknown grids. The map resolution is empirically set based on robot size and environmental complexity, and optimized through simulation verification. In this embodiment, the resolution is [not specified]. The decision-making process adopts an event-driven mechanism, triggering the next round of decision-making when the navigation task is completed or the map is significantly updated. A fallback timer is also set to ensure system responsiveness. To measure the degree of uncertainty in the environment, map entropy is used as the evaluation metric, as shown in formula (1):

[0087]

[0088] in, For grid cells The probability of being occupied. This represents the summation over all raster cells. Map entropy. The larger the size, the greater the degree of uncertainty in the environment.

[0089] In this embodiment, a leading edge point refers to the boundary grid between an unknown region and a known free region on a map. The leading edge candidate point extraction involves the leading edge point detection component 13 checking whether any unknown grid exists in the 8-neighborhood of each free grid; those meeting the criteria are marked as leading edge points. Spatially adjacent leading edge points are clustered into several leading edge regions, and the geometric center of each region is calculated as a candidate point. Leading edge regions that are too small are removed to suppress invalid decisions caused by fragment noise.

[0090] like Figure 2As shown, the structure-aware item calculation component 21 constructs a distance field based on Euclidean distance transformation and designs a structure-aware function to automatically identify topologically critical locations such as narrow connected structures and structural intersections, assigning higher exploration weights to these regions. Key connected structure regions (intersecting diagonal lines) correspond to high structure-aware scores, transitional connected regions (diagonal lines) correspond to medium scores, and open or weakly structured regions correspond to low scores. The distance field is constructed based on Euclidean distance transformation (EDT). As shown in formula (2):

[0091]

[0092] in To occupy the (obstacle) grid set, This indicates taking the minimum value. The structure-aware term first filters candidate points using the bottleneck criterion, and only uses continuous scoring of channel width when conditions such as connectivity and geometric centrality are met; otherwise, it processes the baseline score. Channel width continuous scoring function. It can be written as shown in formula (3):

[0093]

[0094] in The channel width at the candidate point. and These represent the lower and upper limits of the ideal channel width range. For narrow channel threshold, For wide channel threshold, To score for narrow channels, Scoring is given for open or weakly structured areas. Bottleneck criteria include: the point is not close to the map edge, the number of connected components with obstacles within the window is within a reasonable range, the number of local obstacles is moderate, and the estimated channel width falls within the configured bottleneck width range. All of the above thresholds are configurable parameters; this embodiment can use... , , , and take , Baseline score As an example configuration.

[0095] The information gain term calculation is performed by the information gain term calculation component 22, which measures the expected information gain of the robot after reaching the candidate point. This is done on an occupied grid map. (Uncertainty Frontier) is defined as a set of boundary grids that simultaneously satisfy high entropy, high gradient, and are located within a known region, as shown in formula (4):

[0096]

[0097] in For grid The entropy value, Entropy threshold Let be the magnitude of the entropy gradient. For gradient threshold, This is a function to determine whether a grid cell belongs to a known region. The known region refers to grid cells whose probability values ​​are already determined, including occupied grid cells and free grid cells; uncertainty frontier. Typically distributed at the boundary between known and unknown regions, these locations represent potential observation sites where sensors may acquire new information. The entropy gradient is calculated using the Sobel operator; specifically, gradient components are calculated along both the row and column directions of the grid, and the Euclidean norm of both components is taken as the gradient magnitude. Further, the entropy threshold... Can be set to to Any value within the range, as specified in the embodiments of the present invention. Gradient threshold Can be set to to Any value within the range, in this embodiment, is taken as... .

[0098] With candidate points Perform omnidirectional visibility analysis around the center and accumulate visibility. raster entropy, information gain The calculation method is shown in formula (5):

[0099]

[0100] in To start from candidate points Visible Grid set. The final utility is normalized and weighted by the frontier scale, as shown in Equation (6):

[0101]

[0102] in The maximum information gain of all candidate points. The size of the frontier region corresponding to the candidate point. This refers to the frontier scale weighting coefficient. Preferably, the frontier scale weighting coefficient... Can be set to to Any value within the range, in this embodiment, is taken as... .

[0103] The navigation cost term calculation is performed by evaluating the path cost to the candidate point through the navigation cost term calculation component 23. It is defined as a weighted combination of path length, path sparsity, and turning complexity, as shown in formula (7):

[0104]

[0105] in For path length, For path sparsity, For turning complexity, , , These are the corresponding weights. The path sparsity is calculated by statistically analyzing the proportion of obstacle grids within the neighborhood window of each sampling point along the planned path, used to characterize the congestion level of the surrounding environment. The turning complexity is defined as the sum of the absolute values ​​of the turning angles between adjacent road segments on the planned path. Specifically, the path length weight... Can be set to to Any value within the range, in this embodiment, is taken as... Path sparsity weights Can be set to to Any value within the range, as specified in the embodiments of the present invention. Turning complexity weight Can be set to to Any value within the range, in this embodiment, is taken as... When global planning times out or fails, the cost term is approximated using Euclidean distance and the proportion of local obstacles.

[0106] like Figure 4 As shown, the multi-objective utility evaluation of this invention includes three parallel branches: structural feature quantification, information gain assessment, and navigation risk assessment. Finally, a candidate point score is calculated through comprehensive utility calculation. The structural perception term, information gain term, and navigation cost term are weighted and fused to construct a comprehensive utility scoring function, as shown in formula (8):

[0107]

[0108] in For the weights of the structure-aware terms, For the information gain term weight, This refers to the weights of the navigation cost term. This embodiment of the invention employs parameterized weight configuration; a set of optimized configurations in the current implementation is as follows: , , Furthermore, the weights of the structure-aware terms... Can be set to to For any value within the range, the information gain term weight Can be set to to Any value within the range, the weight of the navigation cost term Can be set to to Any value within the range.

[0109] like Figure 5 As shown, a two-stage hierarchical screening strategy based on the Top-K candidate set is adopted to reduce the computational complexity of large-scale frontier point sets. The coarse screening principle is as follows: First, the candidate point set... Perform security checks and filtering to remove unreachable candidate points or those located near obstacles, resulting in a safe candidate set. Then, the path cost is approximated using Euclidean distance to quickly compress the candidate set. To ensure consistency in the scoring scale between the coarse and fine screening, the same absolute benchmark normalization method is used for the Euclidean distance as for the fine screening. The utility score for the coarse screening is shown in formula (9):

[0110]

[0111] in The normalized Euclidean distance approximation cost is calculated as shown in formula (10):

[0112]

[0113] in This is the robot's current position. Candidate points The Euclidean distance to the robot's current position. This is the baseline value for the cost. For Candidate points Sort in descending order and extract the first few truncations. The candidate points are selected for fine screening. The computational complexity of the coarse screening is O(n log n). .

[0114] Fine screening principle: only for Perform real-world planning evaluations on each candidate site. Perform global path planning on each candidate point, obtain the actual path cost, and map it to... to The interval is shown in formula (11):

[0115]

[0116] in For the cost of planning the path, This is the cost benchmark value. To maintain consistency with the coarse screening, the cost benchmark value in this embodiment of the invention is taken as... Overall, the cost baseline can be set according to the environmental complexity. to Any value within the range. The final utility score of the refined screening has the same form as Formula 8, the difference being the cost term. Adopting real planning costs As shown in formula (12):

[0117]

[0118] The candidate point with the highest utility score is selected as the optimal target point, as shown in formula (13):

[0119]

[0120] The computational complexity of fine screening is ,in This represents the computational complexity of a single path planning operation. The total computational complexity is... This is significantly lower than performing exact global planning on all candidate points. Complexity. Number of Top-K selections Can be set to to Any integer within the range, as specified in the embodiments of the present invention. To address the real-time constraints of computing-limited platforms, a time limit is imposed on each planning task, employing... (Request timed out) and The two-level timeout mechanism (result waiting timeout) degrades the cost term of the corresponding candidate point to an approximation of Euclidean distance and local congestion when the planning timeout or fails.

[0121] The global path planning component 31 is used to perform global path planning on the selected optimal target point, generating a global path from the robot's current position to the target point. This embodiment of the invention uses the SmacPlanner2D planner of the Nav2 navigation stack. This planner is based on a hybrid A* algorithm and has better dynamic map adaptability compared to traditional potential field planners, effectively handling scenarios with frequent map updates during SLAM exploration. The local trajectory tracking component 32 is used to perform local trajectory tracking. This embodiment uses the DWB (Dynamic Window Based) local planner of the Nav2 navigation stack, generating a smooth local trajectory based on the global path and real-time perception information. The motion control component 41 is used to control the robot's motion. This embodiment uses a differential drive control model, controlling the robot's motion based on the linear velocity and angular velocity commands output by the local planner. The state feedback component 42 is used to provide real-time feedback of the robot's state information, including position, velocity, and heading, providing real-time pose information to the perception module 1 through TF coordinate transformation.

[0122] To improve the robustness of exploration completion determination, this embodiment of the invention employs a continuous counting mechanism: exploration is considered complete only when no valid leading edge points are detected consecutively multiple times. Specifically, the continuous counting threshold can be set to any integer within the range of 2 to 5 times; in this embodiment, it is set to 2 times. When the candidate point set is empty, the robot's current position is first checked to see if it is near a topologically critical region (such as a narrow connected structure or a structural intersection). The topologically critical region is identified through distance field analysis; it is determined to be a topologically critical region when the channel width at the detection point falls within a preset bottleneck width range. If a topologically critical region is detected, a map update is actively triggered, and a re-check is performed after a preset delay, preferably 0.5 seconds, to wait for the SLAM system to complete the map update. If no topologically critical region is detected, the continuous no-leading-edge counter is incremented. This active re-check mechanism effectively avoids misjudgments caused by map update delays or sensor blind spots, improving the exploration completion rate.

[0123] like Figures 3-6 As shown, the present invention provides a robot autonomous exploration and mapping method based on structure perception and multi-objective decision-making, comprising the following steps:

[0124] Step S1: Environmental characterization and map construction;

[0125] The LiDAR sensor 11 collects 360° distance information of the environment, and the raster map building component 12 generates a resolution of [resolution missing]. The occupied grid map is calculated, and the map entropy is calculated according to formula (1). To measure the degree of uncertainty in the environment.

[0126] Step S2: Extraction of frontier candidate points;

[0127] The leading edge detection component 13 is invoked to detect the unknown-free region boundary grid on the grid map, cluster the leading edge points and calculate the geometric center of each region to generate a candidate point set, while removing excessively small leading edge regions to suppress noise.

[0128] Step S3: Calculation of multi-objective utility indicators;

[0129] The following three sub-steps are performed in parallel on the candidate point set:

[0130] Sub-step S31 (structural feature quantization): Call the structure-aware item calculation component 21, obtain the distance field value at each candidate point according to formula (2), and calculate the structure-aware score according to formula (3) after bottleneck criterion screening. .

[0131] Sub-step S32 (Information Gain Assessment): Call the information gain calculation component 22, construct the UF set according to formula (4), perform omnidirectional visibility analysis with candidate points as the center, calculate the information gain according to formula (5), and then normalize it according to formula (6) to obtain the result. .

[0132] Sub-step S33 (Navigation Risk Assessment): Call the navigation cost calculation component 23 to calculate the path cost according to formula (7). If the global planning times out or fails, an approximation is made using Euclidean distance and the proportion of local obstacles.

[0133] Step S4: Construction of a multi-objective utility model;

[0134] According to formula (8), substitute the calculation results of S31, S32, and S33 into the formula to calculate the comprehensive utility score of each candidate point. .

[0135] Step S5: Top-K two-stage stratified screening;

[0136] Calling the Top-K hierarchical filtering component 24:

[0137] Sub-step S51 (coarse screening): After performing security checks and filtering on the candidate point set, calculate the coarse screening utility score according to formula (9) and formula (10) with Euclidean distance approximation cost, and extract the top K candidate points in descending order.

[0138] Sub-step S52 (Fine Screening): For the K candidate points, call the global path planning component 31 to obtain the actual path cost, normalize according to formula (11) and enter it into formula (12) to calculate the fine screening utility score, and select the one with the highest utility. As the optimal target point, as shown in formula (13).

[0139] Step S6: Target selection and motion execution;

[0140] The optimal target point The data is transmitted to the planning module 3, where the global path planning component 31 is called to generate a global path. The local trajectory tracking component 32 generates a smooth local trajectory, and the motion control component 41 controls the robot to perform motion according to the differential drive model. The state feedback component 42 feeds back the pose information to the perception module 1 in real time through TF coordinate transformation.

[0141] Step S7: Candidate point set empty determination and loop control;

[0142] Perform front-edge detection on the grid map: (1) If the candidate point set is not empty, reset the consecutive no-front-edge counter and return to step S1 to continue exploration; (2) If it is empty, check whether the robot's current position is near the topological critical region. If so, actively trigger map update and recheck after a 0.5-second delay; otherwise, increment the consecutive no-front-edge counter; (3) When the consecutive empty detection reaches 2 times, determine that the exploration is complete and end the process. This step realizes closed-loop control of the exploration process, ensuring that the robot can autonomously and completely complete the environmental mapping task.

[0143] In multi-type structured simulation environments that include channels, regions, and bottlenecks, the path intersections and overlapping coverage of the method of this invention are significantly reduced, and the mapping process is more stable. Figure 6 (a): This illustrates the typical exploration trajectory of the method of the present invention, which generally presents a behavior pattern of first advancing the corridor connection and then entering the room to complete the coverage. This is consistent with the structural perception item's preference for doorways and key connection locations. Figure 6 (b): This demonstrates the comparative trajectory of the nearest frontier strategy, exhibiting characteristics of local optima and numerous redundant paths and efficiency losses. Under the same simulation conditions, experimental results further show that: compared to the nearest distance strategy, the exploration completion time of this invention is reduced by 16.4%, and the path length is reduced by 14.3%; compared to the fixed weight strategy, the exploration completion time is reduced by 26.6%, and the path length is reduced by 29.3%. Furthermore, the coefficient of variation of this invention is less than 5% in the later stages of exploration, demonstrating good decision stability.

Claims

1. A robot autonomous exploration and mapping system, characterized in that: The system includes a perception module (1), a decision-making module (2), a planning module (3), and a control module (4). The output of the perception module (1) is connected to the input of the decision-making module (2). The perception module (1) transmits the candidate set of leading edge points to the decision-making module (2). The output of the decision-making module (2) is connected to the input of the planning module (3). The decision-making module (2) transmits the optimal target point to the planning module (3). The output of the planning module (3) is connected to the input of the control module (4). The planning module (3) transmits motion commands to the control module (4). The output of the control module (4) is connected to the input of the perception module (1). The control module (4) feeds back the state information to the perception module (1), forming a closed-loop control system.

2. The robot autonomous exploration mapping system according to claim 1, characterized in that: The perception module (1) includes a LiDAR sensor (11), a grid map building component (12), and a leading edge detection component (13). The LiDAR sensor (11) is used to collect environmental distance information. The grid map building component (12) is connected to the LiDAR sensor (11) to convert the distance information into an occupied grid map and calculate the map entropy to measure the degree of environmental unknown. The leading edge detection component (13) is connected to the grid map building component (12) to detect the boundary grid between the unknown area and the known free area. It generates a set of candidate points through clustering and transmits the information to the decision module (2).

3. The robot autonomous exploration mapping system according to claim 1, characterized in that: The decision module (2) includes a structure perception item calculation component (21), an information gain item calculation component (22), a navigation cost item calculation component (23), and a Top-K hierarchical screening component (24). The structure perception item calculation component (21), the information gain item calculation component (22), and the navigation cost item calculation component (23) transmit the calculation results to the Top-K hierarchical screening component (24). The Top-K hierarchical screening component (24) performs a two-stage screening of coarse screening and fine screening based on the received three scores, and outputs the optimal target point to the planning module (3).

4. The robot autonomous exploration mapping system according to claim 1, characterized in that: The planning module (3) includes a global path planning component (31) and a local trajectory tracking component (32) for signal connection. The global path planning component (31) performs path planning on candidate points to obtain the true cost; the local trajectory tracking component (32) generates a smooth local trajectory based on the global path.

5. The robot autonomous exploration mapping system according to claim 1, characterized in that: The control module (4) includes a motion control component (41) and a state feedback component (42). The motion control component (41) is connected to the planning module (3) by signal, and the state feedback component (42) transmits the robot pose information to the perception module (1).

6. A method for performing autonomous exploration mapping of a robot using the robot autonomous exploration mapping system according to any one of claims 1 to 5, characterized in that... Includes the following steps: Step S1: Environmental characterization and map construction; The environmental information is collected by the sensing module (1), an occupied grid map is generated, and the map entropy H(M) is calculated to measure the degree of environmental unknown. Step S2: Extraction of frontier candidate points; The perception module (1) detects the unknown-free region boundary grid on the grid map, clusters the leading points and calculates the geometric center of each region, generates a candidate point set, and removes leading regions with fewer than 3 leading points to suppress noise. Step S3: Calculation of multi-objective utility indicators; The decision module (2) obtains the distance field at each candidate point and calculates the structure perception score. ; The decision module (2) constructs a UF set on the occupied grid map, performs omnidirectional visibility analysis with candidate points as the center, calculates information gain, and then normalizes to obtain the frontier scale weight. ; Decision module (2) calculates path cost ; If global planning times out or fails, an approximation is made using Euclidean distance and the proportion of local obstacles. Step S4: Construction of a multi-objective utility model; According to the obtained , , Calculate the overall utility score for each candidate point. ; Step S5: Top-K two-stage stratified screening; After performing security checks and filtering on the candidate point set, the decision module (2) calculates the coarse screening utility score using Euclidean distance as an approximation cost, and sorts the top K candidate points in descending order; the planning module (3) obtains the actual path cost for the K candidate points, calculates the fine screening utility score, and selects the one with the highest utility. As the optimal target point; Step S6: Target selection and motion execution; The optimal target point The data is transmitted to the planning module (3) to generate a global path and a smooth local trajectory. The control module (4) controls the robot to perform motion according to the differential drive model and feeds back the pose information to the perception module (1) in real time. Step S7: Candidate point set empty determination and loop control; Perform front-edge detection on the grid map. If the candidate point set is not empty, reset the consecutive no-front-edge counter and return to step S1 to continue exploration. If the candidate point set is empty, check whether the robot's current position is near the topological critical region. If so, actively trigger map update and recheck after a 0.5-second delay. Otherwise, increment the consecutive no-front-edge counter. When two consecutive empty checks are performed, the exploration is considered complete and the process ends.

7. A robot autonomous exploration mapping method according to claim 6, characterized in that, In step S1, the formula for calculating H(M) is: ； in, For grid cells The probability of being occupied. This represents the summation over all raster cells, specifically map entropy. The larger the value, the higher the degree of uncertainty in the environment.

8. The robot autonomous exploration mapping method according to claim 1, characterized in that, In step S3, a range field is constructed based on the Euclidean distance transform (EDT). The mathematical model is: ； in For the set of occupied / obstacle grids, This indicates that the minimum value is taken; the structure-aware item first filters candidate points through the bottleneck criterion, and only uses continuous scoring of channel width when conditions such as connectivity and geometric centrality are met; otherwise, it is processed according to the baseline score. Channel width continuous scoring function for: ； in The channel width at the candidate point. and These represent the lower and upper limits of the ideal channel width range. For narrow channel threshold, For wide channel threshold, To score for narrow channels, Scoring is given for open or weakly structured areas; Bottleneck criteria include: the point is not close to the map edge, the number of obstacles within the window is within a reasonable range, the number of local obstacles is moderate, and the estimated channel width falls within the configured bottleneck width range. The UF set is a set of boundary grids that simultaneously satisfy high entropy, high gradient, and are located within a known region, denoted as: ； in For grid The entropy value, Entropy threshold Let be the magnitude of the entropy gradient. For gradient threshold, A function to determine whether a grid belongs to a known region; The known region refers to a grid cell with a determined probability value, including occupied grid cells and free grid cells; uncertainty frontier They are typically located at the boundary between known and unknown areas, representing potential observation locations where sensors may acquire new information; With candidate points Perform omnidirectional visibility analysis around the center and accumulate visibility. raster entropy, information gain The representation is: ； in To start from candidate points Visible The grid set; the final utility, after normalization and the introduction of frontier scale weights, is expressed as: ； in The maximum information gain of all candidate points. The size of the frontier region corresponding to the candidate point. For frontier size weighting coefficients; The path cost is a weighted combination of path length, path sparsity, and turning complexity, expressed as: ； in For path length, For path sparsity, For turning complexity, , , These are the corresponding weights; the path sparsity is calculated by statistically analyzing the proportion of obstacle grids within the neighborhood window of each sampling point on the planned path, and is used to characterize the congestion level of the surrounding environment of the path; the turning complexity is defined as the sum of the absolute values ​​of the turning angles between adjacent road segments on the planned path.

9. A robot autonomous exploration mapping method according to claim 1, characterized in that, In step S4, the multi-objective utility evaluation includes three parallel branches: structural feature quantification, information gain assessment, and navigation risk assessment. Finally, a candidate point score is calculated through comprehensive utility calculation. The structural perception term, information gain term, and navigation cost term are weighted and fused to construct a comprehensive utility scoring function, expressed as: ； in For the weights of the structure-aware terms, For the information gain term weight, This refers to the weight of the navigation cost term.

10. A robot autonomous exploration mapping method according to claim 1, characterized in that, In step S5, the two-stage stratified screening consists of coarse screening and fine screening. The coarse screening first examines the candidate point set. Perform security checks and filtering to remove unreachable candidate points or those located near obstacles, resulting in a safe candidate set. Then, the path cost is approximated using Euclidean distance to quickly compress the candidate set, and the Euclidean distance is normalized using an absolute benchmark; the utility score of the coarse screening is expressed as: ； in The cost of the normalized Euclidean distance approximation is expressed as: ； in This is the robot's current position. Candidate points The Euclidean distance to the robot's current position. As the baseline value for cost; for Candidate points Sort in descending order and extract the first few truncations. Each candidate point enters the fine screening; the computational complexity of the coarse screening is O(n). ; Fine screening is only for Perform real-world planning assessments on each candidate site. Perform global path planning on each candidate point, obtain the actual path cost, and map it to... to The interval, the true path cost is represented as: ； in For the cost of planning the path, The cost baseline value; the final utility score of the refined screening is expressed as: ； Among them, the cost item Adopting real planning costs As shown in formula (12): Select the candidate point with the highest utility score as the optimal target point, denoted as: ； The computational complexity of fine screening is ,in The computational complexity of a single path planning operation is ; the total computational complexity is . This is lower than performing exact global planning on all candidate points. Complexity; Number of Top-K selections for to Any integer within the range.

Images