A large model driven path growth type semantic topology self-optimization navigation decision loop method in an unknown environment

Abstract

The application discloses a large model driven path growth type semantic topology self-optimization navigation decision closed loop method in an unknown environment, takes the traveled path of a robot as the only topology skeleton, incrementally constructs a path growth type semantic topology map, only includes the path nodes of the traveled area, the surrounding task related semantic nodes and the front exploration nodes, does not store irrelevant environment information, and completely eliminates redundancy from the source; in combination with large model driven multi-dimensional self-optimization operation, the topology structure is further simplified, the map compactness is maintained, the topology map storage and the calculation resource occupation are reduced by more than 50%, the nodes and edges are subjected to differential multi-dimensional feature coding, the node coding space, semantic, task and state information are coded, sufficient feature information is ensured, excessive redundancy is avoided, sufficient and accurate judgment bases are provided for topology self-optimization and navigation decision, in combination with navigation decision, only the optimized effective feature information is called, the decision calculation amount is greatly reduced, and the planning efficiency is improved by more than 20%.

Description

Technical Field

[0001] This invention relates to the field of autonomous robot navigation technology, specifically to a path-growing semantic topology self-optimizing navigation decision-making closed-loop method driven by a large model in unknown environments, applicable to target arrival navigation tasks for various mobile robots such as wheeled and legged robots in complex and unknown indoor and outdoor environments. Background Technology

[0002] Autonomous navigation in unknown environments is a core technology for mobile robots, widely used in practical scenarios such as home services, elderly care services, emergency rescue, target search, park inspection, and field exploration. As the complexity and scale of application environments increase, higher demands are placed on robots' large-scale environmental perception and long-field-of-view decision-making and planning capabilities. Due to its lightweight and structured characteristics, topological maps have become the mainstream map representation method to replace traditional dense quantitative maps.

[0003] Traditional global topology map construction methods indiscriminately incorporate all environmental information perceived by the robot into the topology structure. This leads to the topology map growing infinitely as exploration progresses, resulting in severe information redundancy. This not only consumes significant storage and computing resources but also forces subsequent navigation decisions to process a large amount of irrelevant information, drastically reducing decision-making efficiency. To address the problem of weak semantic representation, existing technologies introduce large models for semantic information processing. However, these large models are limited to the parsing and classification of single topology nodes, neglecting their core advantages in information association, organization, and reasoning. They are not applied to the integration of topology nodes themselves or the optimization of the topology structure, resulting in a severe disconnect between semantic processing and topology optimization. This fails to fundamentally solve the problem of topology information redundancy and falls far short of the human mind's ability to associate, organize, and reason about environmental information.

[0004] Meanwhile, in existing technologies, the relationship between the topology map and navigation decisions is only one-way; the topology map provides input for the decision, but information such as topological deviations and environmental changes discovered during the decision-making process cannot guide the optimization of the topology map in the reverse direction, resulting in extremely low coupling between the two. When the decision-making process discovers that the node / edge features of the topology map do not match the actual environment, or that new traversable paths exist, the topology map cannot be corrected in time, which can easily lead to subsequent planning errors, redundant movements, getting stuck in local optima, or even failing to reach the target. This is also an important reason for the performance degradation of traditional global topology navigation in long-sequence tasks in complex and unknown environments.

[0005] Furthermore, traditional topology maps use a single, undifferentiated method for encoding the features of nodes and edges, or lack sufficient feature information to provide effective criteria for topology optimization and pruning operations. Alternatively, excessive feature information can further increase the computational burden, making it difficult to balance the core requirements of "sufficiency of information" and "non-redundancy," and failing to provide convenient and accurate criteria for subsequent topology processing.

[0006] In summary, existing robot navigation technologies for unknown environments have the following problems:

[0007] 1) The global topology map has information redundancy, which can easily lead to ineffective expansion and affect maintenance efficiency and navigation efficiency.

[0008] 2) The combination of large models and navigation strategies has limitations in application, failing to provide effective support for topology optimization and decision planning, and failing to fully realize the value of large models and topology.

[0009] 3) The low coupling between topology and navigation decision-making leads to low overall navigation efficiency, poor environmental adaptability, difficulty in achieving dynamic collaborative iteration, and inability to meet the long-field-of-view navigation requirements in complex and unknown environments. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a closed-loop navigation and decision-making method for path growth-based semantic topology self-optimization driven by a large model in unknown environments. This method solves the technical problems of redundant global topology map information, single mode of large model information processing and combination, and low coupling between topology and decision-making in traditional technologies. It achieves lightweight construction of topology maps, deep empowerment of large models, and bidirectional coupling between topology and decision-making, thereby improving the navigation efficiency and environmental adaptability of robots in complex and unknown environments.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] This invention discloses a path-growing semantic topology self-optimizing navigation decision-making closed-loop method driven by a large model in unknown environments, comprising:

[0013] The robot's motion state information and environmental perception information are collected through multi-source sensing devices;

[0014] Based on the robot's motion state information, the already traveled path is solved as the skeleton, and nodes and edges are incrementally generated to obtain a path-growing semantic topology map.

[0015] Multimodal large models are used to encode, infer, and complete environmental perception information in multiple dimensions, and to obtain the node and edge information annotation and feature encoding of path-growing semantic topology map.

[0016] Based on the annotation and feature encoding of node and edge information of the path-growing semantic topology map, spatial connectivity and semantic relevance are determined, and the structure of the path-growing semantic topology map is optimized to obtain a path-growing semantic topology map with self-optimized structure.

[0017] Based on the path-growing semantic topology map after structural self-optimization, a hierarchical navigation strategy is adopted to make decisions, output navigation control commands, and obtain navigation execution feedback based on the execution status.

[0018] Based on navigation execution feedback, the task completion status is determined. If the goal is not reached, the task is retried; if the goal is reached, the task ends, forming a closed-loop iteration that drives the dynamic correction of the topology map structure.

[0019] As a further improvement, the multi-source sensing device of this invention collects robot motion state information and environmental perception information. Specifically, the multi-source sensing device is compatible with LiDAR, visual sensors, IMU, and odometry, and does not require the pre-construction of a global dense map. The collected data supports online topology construction, optimization, and navigation decisions throughout the entire process. The robot motion state information is used to solve the robot's already traveled path, providing a foundation for the skeleton construction of a path-growing semantic topology map. The environmental perception information is used for differentiated multi-dimensional encoding of multimodal large models, providing data support for the information annotation, feature encoding, and topology self-optimization of nodes and edges in the path-growing semantic topology map. The robot motion state information and environmental perception information are subsequently used for the construction of the path-growing semantic topology map.

[0020] As a further improvement, the path-growing semantic topology map of the present invention is specifically constructed by incrementally generating topology map nodes and edges using the robot's already traveled path as the unique skeleton. During the construction process, only the path nodes of the already traveled area, the semantic nodes related to the surrounding tasks, and the frontier exploration nodes are included, and irrelevant environmental information that is not related to the already traveled path is not stored.

[0021] As a further improvement, the node and edge information annotation and feature encoding of the path-growing semantic topology map of this invention are implemented through a multimodal large model based on environmental perception information. Specifically, the multimodal large model performs heterogeneous data parsing, reasoning, and prior completion on the environmental perception information, outputs semantic analysis conclusions, and completes the information annotation and differentiated multidimensional feature encoding of nodes and edges accordingly. During this process, the large model simultaneously completes the unified parsing of heterogeneous semantic data, fuzzy information reasoning, local semantic completion, as well as task relevance scoring and feature encoding verification and completion.

[0022] As a further improvement, the path-growing semantic topology map of the present invention is obtained based on the path-growing semantic topology map and its node and edge information annotation and feature encoding. Specifically, by combining the feature encoding information of nodes and edges, spatial connectivity and semantic relevance are judged, and a multimodal large model performs topology self-optimization operation in real time to realize dynamic optimization of the initially established path-growing semantic topology map, thereby obtaining the path-growing semantic topology map after structural self-optimization.

[0023] As a further improvement, the topology self-optimization operation of this invention specifically includes: topology node integration, intelligent pruning, branch merging, cross-branch shorting, dynamic node addition and deletion, and branch weight adjustment. Among them, intelligent pruning removes invalid and redundant branches, branch merging simplifies similar nodes and adjacent edges, cross-branch shorting connects non-adjacent nodes through reasoning, dynamic node addition and deletion adapts to environmental changes, and branch weight adjustment allocates exploration priority according to task value. Finally, a compact, efficient, and adaptable self-optimized topology map that fits the current environment and navigation task is obtained.

[0024] As a further improvement, the navigation execution feedback of this invention is based on a path-growing semantic topology map after structural self-optimization. Specifically, based on the structurally self-optimized topology map, a three-level hierarchical navigation strategy of high-level frontier exploration node selection, mid-level path planning, and low-level dynamic obstacle avoidance is adopted for decision-making, and navigation control commands including topology node selection and speed control are output. By collecting the execution status and results of the navigation control commands, navigation execution feedback is generated, which includes planning and obstacle avoidance results, frontier node selection status and call frequency, environmental change data, and motion execution deviation information.

[0025] As a further improvement, the navigation execution feedback of this invention is used to drive the dynamic correction of the topology map structure. Specifically, the planning and obstacle avoidance results in the navigation execution feedback are fed back to the topology self-optimization stage in real time to update the feature encoding of topology nodes and edges, mark invalid connected edges, add cross-branch short-circuit paths, adjust the dynamic obstacle risk of nodes and the passage cost of edges, and trigger the dynamic correction of the topology map. The selection status and call frequency of the leading edge nodes are fed back synchronously to update the task characteristics and state characteristics of the nodes, so as to realize the dynamic matching between the topology map and the navigation decision requirements. At the same time, the task execution status is judged according to the navigation execution feedback. If the navigation target is not reached, the entire navigation process is iterated in a loop. If the target is reached, the task ends, realizing the closed-loop operation of the topology-decision bidirectional coupling.

[0026] As a further improvement, the method of this invention is decoupled from the robot hardware platform and the type of multi-source sensing device, and can be adapted to various mobile robots, including wheeled and tracked robots. It does not require the pre-construction of a global dense map. Relying on the robot's motion state information and environmental perception information, it completes the construction of a path-growing semantic topology map, the annotation and feature encoding of path-growing semantic topology map nodes and edges, the generation of a path-growing semantic topology map after structural self-optimization, the hierarchical navigation strategy decision-making, and the navigation execution feedback loop online. It can effectively adapt to unknown environments in multiple indoor and outdoor scenarios and meet the long-field-of-view navigation requirements in complex and unknown environments.

[0027] Beneficial effects

[0028] This invention discloses a path-growing semantic topology self-optimizing navigation decision-making closed-loop method and system driven by a large model in unknown environments. It relates to the field of robot autonomous navigation technology and specifically addresses the core technical pain points of robot navigation in unknown environments. Compared with existing technologies, the specific beneficial effects brought by each innovation are as follows:

[0029] (1) By using a path-growing semantic topology map, the technical pain point of information redundancy in traditional global topology maps is solved: Existing technologies use global topology maps to store all environmental information indiscriminately, which easily leads to meaningless expansion of the topology, excessive storage and computing resource consumption, and cannot adapt to the limited computing power of mobile robots. This invention uses the robot's already traveled path as the only topology skeleton, incrementally constructs a path-growing semantic topology map, and only includes path nodes in the already traveled area, semantic nodes related to the surrounding tasks, and nodes for frontier exploration. It does not store irrelevant environmental information, thus eliminating redundancy from the root. Combined with a large model-driven multi-dimensional self-optimization operation, the topology structure is further simplified and the map is kept compact. Ultimately, the storage and computing resource consumption of the topology map is reduced by more than 50%, which can adapt to mobile robots with various hardware computing power limitations such as wheeled and tracked robots, and realize the lightweight construction of the topology map.

[0030] (2) By using path-growing semantic topology map node and edge information annotation and feature encoding, the problem of unreasonable traditional topology feature encoding is solved: In the existing technology, the feature encoding of topology nodes and edges lacks differentiated design, either the information is insufficient to support optimization and decision-making, or the information is redundant and increases the computational burden. This invention performs differentiated multi-dimensional feature encoding on nodes and edges. Nodes encode spatial, semantic, task and state information, while edges encode connectivity, cost, type and state information. This ensures sufficient feature information while avoiding excessive redundancy, providing sufficient, convenient and accurate judgment basis for topology self-optimization and navigation decision-making. At the same time, combined with navigation decision-making, only the optimized effective feature information is called, which greatly reduces the amount of decision-making computation and improves planning efficiency by more than 20%.

[0031] (3) By performing structural self-optimization operations, the limitations of large models in topology navigation are overcome: In existing technologies, large models are only used for single-node information processing and cannot participate in topology construction and structural optimization, resulting in low utilization of semantic information and poor adaptability of topology structure. This invention relies on large models to not only complete heterogeneous semantic data parsing, fuzzy information reasoning, local semantic prior completion, task relevance scoring and feature encoding verification and completion in the topology establishment stage, but also realize multi-dimensional self-optimization such as topology node integration, intelligent pruning, and branch merging. It simulates the human ability to associate, organize and complete information, realizes the deep integration of semantic processing and topology optimization, avoids the loss of effective semantic information caused by simple spatial logic processing, significantly improves the topology map's ability to represent complex and unknown environments, and further optimizes storage and resource consumption, giving full play to the empowering value of large models.

[0032] (4) By implementing navigation execution feedback, a two-way coupled closed loop of topology and decision-making is achieved, breaking through the limitations of traditional one-way drive: In the existing technology, topology construction and navigation decision-making are independent and driven in one direction. The topology map cannot be dynamically corrected according to decision feedback, which easily leads to planning deviations and redundant movements. This invention constructs a three-level hierarchical navigation strategy and, through a reverse feedback mechanism, feeds back the low-level path planning, local dynamic obstacle avoidance results, and front-end node selection to the topology self-optimization stage in real time, driving the dynamic correction of the topology map and the update of feature encoding. This achieves dynamic collaborative iteration of topology construction and navigation decision-making, greatly improving the accuracy of navigation planning, effectively avoiding redundant robot movements caused by planning errors, and improving navigation smoothness.

[0033] (5) Based on the synergistic effect of all the above innovations, the navigation efficiency and target achievement success rate are greatly improved: Compared with the existing technology, in complex and unknown environments, the present invention can reduce the redundant motion of the robot by more than 20%, achieve a target achievement success rate of up to 80%, and increase the path length weighted success rate (SPL) by 15%-25%. It can significantly solve the problems of low navigation efficiency and poor environmental adaptability of traditional technologies, and significantly improve the stability and reliability of long field of view navigation in complex and unknown environments, so as to meet the actual navigation needs of various mobile robots. Attached Figure Description

[0034] Figure 1 This is a block diagram of the overall architecture of the system of the present invention;

[0035] Figure 2 This is a flowchart illustrating the method of the present invention for searching for missing persons in campus security. Detailed Implementation

[0036] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0037] This invention discloses a path-growing semantic topology self-optimizing navigation decision-making closed-loop method driven by a large model in unknown environments. It can be applied to campus security missing persons search tasks. The implementing entity is a wheeled campus security patrol robot equipped with a multi-source sensing device. The task scenario is an unknown area of ​​the campus without a pre-built global dense map (including newly built teaching areas not yet entered into the system, the area around temporarily controlled sports venues, building corridors and underground parking garage entrance areas). Figure 1 This is a block diagram of the overall architecture of the system of the present invention; Figure 2 This is a flowchart illustrating the method for searching for missing persons in campus security, as described in this invention. The technical solution of this invention will be further described below with reference to the accompanying drawings and specific embodiments:

[0038] The specific implementation is as follows:

[0039] 1) Start: The robot is powered on and enters a ready-to-go state;

[0040] 2) Collect robot motion state information and environmental perception information through multi-source sensing devices. Specifically, the multi-source sensing devices collect motion data such as robot pose and speed during the campus search process in real time, as well as environmental data such as campus roads, obstacles, personnel characteristics, and area markers, to provide basic input for the entire process.

[0041] The multi-source sensing device collects robot motion state information and environmental perception information. Specifically, the multi-source sensing device is compatible with LiDAR, vision sensors, IMU, and odometry. It does not require the pre-construction of a global dense map, and the collected data supports online topology construction, optimization, and navigation decisions throughout the entire process. In this scenario, the multi-source sensing device specifically adopts a 16-line LiDAR, a binocular visible light / infrared vision sensor, a 9-axis IMU, and a wheeled odometry. It does not require the pre-construction of a global campus map, and all collected data is processed online throughout the entire process, supporting topology construction, optimization, and navigation decisions throughout the campus search process.

[0042] Robot motion state information is used to solve the robot's traveled path, providing a foundation for the skeleton construction of a path-growing semantic topology map. In this scenario, the robot motion state information specifically includes the robot's real-time pose coordinates, heading angle, travel speed, cumulative mileage, and turning angle data, which are used to calculate the robot's traveled path within the campus, providing a core foundation for the topology map skeleton construction.

[0043] Environmental perception information is used for differentiated multi-dimensional encoding of multimodal large models, providing data support for information annotation, feature encoding, and topology self-optimization of nodes and edges in path-growing semantic topology maps. Robot motion state information and environmental perception information are subsequently used to construct path-growing semantic topology maps. In this scenario, environmental perception information specifically includes 3D point cloud data of campus roads, building walls, temporary fences, green belts, and obstacles collected by LiDAR, semantic data of campus road signs, building signs, personnel features, and road surface status images collected by visual sensors, as well as real-time monitoring data of dynamic obstacles and ambient lighting. This provides core data support for large model encoding, node feature annotation, and topology optimization. Both types of information are synchronously transmitted to the topology map construction stage.

[0044] 3) Based on the robot's motion state information, solve the already traveled path as the skeleton, incrementally generate nodes and edges, and obtain a path-growing semantic topology map; specifically, take the actual travel path of the robot after starting from the security booth as the unique skeleton, and incrementally generate topology nodes and edges as the robot searches, and construct a path-growing semantic topology map adapted to the campus search scenario.

[0045] The path-growing semantic topology map is constructed by incrementally generating topology map nodes and edges, using the robot's already traveled path as the sole skeleton. During construction, only nodes from the already traveled area, task-related semantic nodes in the surrounding area, and nodes for forward exploration are included; irrelevant environmental information unrelated to the already traveled path is not stored. In this scenario, the topology map construction process only includes three types of core nodes: first, nodes from the already traveled area, i.e., pose nodes set every 10 meters along the robot's already traveled path on campus; second, task-related semantic nodes in the surrounding area, i.e., nodes within 10 meters of the already traveled path that are strongly related to personnel search, such as building entrances / exits, road intersections, open areas, corridor entrances, and areas around benches where people are likely to linger; and third, nodes for forward exploration, i.e., unexplored road intersections and accessible area entrances at the end of the already traveled path. Environmental information unrelated to search, such as the interior of green belts outside the already traveled path, the back of buildings, and enclosed logistics areas, is not included in the map storage, thus avoiding meaningless topology expansion from the outset.

[0046] 4) Utilize multimodal large models to encode, reason, and complete environmental perception information in multiple dimensions, and obtain path-growing semantic topology map node and edge information annotation and feature encoding; specifically, use pre-trained multimodal large models to parse, reason, and complete the collected heterogeneous campus environment data, and complete the topology node and edge information annotation and feature encoding to meet the task requirements of personnel search.

[0047] The path-growing semantic topology map node and edge information annotation and feature encoding are implemented through a multimodal large model based on environmental perception information. Specifically, the multimodal large model performs heterogeneous data parsing, inference, and prior completion on environmental perception information, outputting semantic analysis conclusions, and completing node and edge information annotation and differentiated multi-dimensional feature encoding accordingly. During this process, the large model simultaneously completes unified parsing of heterogeneous semantic data, fuzzy information inference, local semantic completion, task relevance scoring, and feature encoding verification and completion. In this scenario, a pre-trained multimodal large model is used to uniformly parse heterogeneous campus environmental data such as laser point clouds and visual images, infer and judge fuzzy semantic information caused by nighttime occlusion and insufficient light, and complete prior knowledge of campus scenes in perception blind spots (such as building entrances and exits necessarily connecting to internal roads, and green belts being normally impassable). At the same time, based on the personnel search task, the task relevance and personnel dwell probability of nodes / edges are quantitatively scored, and semantic analysis conclusions are output, thereby completing node and edge information annotation and differentiated multi-dimensional feature encoding. The correlation verification and completion of feature encoding are completed simultaneously to ensure information integrity and consistency.

[0048] 5) Based on the node and edge information annotation and feature encoding of the path-growing semantic topology map, determine the spatial connectivity and semantic relevance, and perform structural optimization on the path-growing semantic topology map to obtain a self-optimized path-growing semantic topology map; specifically, combine the spatial connectivity of the campus scene and the semantic relevance of the search task to perform structural self-optimization on the topology map to obtain a self-optimized topology map adapted to the campus environment search.

[0049] The self-optimized path-growing semantic topology map is obtained based on the path-growing semantic topology map and its node and edge information annotations and feature encodings. Specifically, by combining the feature encoding information of nodes and edges, spatial connectivity and semantic relevance are judged, and a multimodal large model performs a real-time structural self-optimization operation to dynamically optimize the initially established path-growing semantic topology map, resulting in a self-optimized path-growing semantic topology map. In this scenario, by combining the spatial connectivity of the campus scene (such as whether roads are passable and whether areas are connected) with the semantic relevance of the personnel search task (such as whether the area is a place where people are likely to linger and whether it is related to the search target), a multimodal large model performs a real-time structural self-optimization operation to dynamically optimize the initially constructed campus scene topology map, ultimately obtaining a self-optimized topology map adapted to the search of unknown campus environments.

[0050] The structural self-optimization operations specifically include: topology node integration, intelligent pruning, branch merging, cross-branch shorting, dynamic node addition and deletion, and branch weight adjustment. Intelligent pruning eliminates invalid and redundant branches; branch merging simplifies similar nodes and adjacent edges; cross-branch shorting connects non-adjacent nodes through reasoning; dynamic node addition and deletion adapt to environmental changes; and branch weight adjustment allocates exploration priority based on task value. Ultimately, a compact, efficient, and adaptable self-optimized topology map for the current environment and navigation task is obtained. In this scenario, the specific implementation of each self-optimization operation is as follows: topology node integration merges adjacent path nodes with consistent semantic attributes at the same intersection on campus; intelligent pruning eliminates invalid and redundant branches such as dead ends, closed construction areas, high-cost stair paths, and high-risk side roads with high pedestrian traffic. The system employs several techniques: First, redundant nodes that are repeatedly detected are removed. Second, branch merging is used to merge adjacent edges with similar connectivity and travel costs within the campus, simplifying the topology. Third, cross-branch short-connection is performed, where the large model, based on prior knowledge of the campus, determines that two non-adjacent building corridor entrances are semantically connected and passable, and then performs cross-branch short-connection and adds edge features. Fourth, dynamic node addition and deletion are implemented, automatically adding nodes when unexplored underground parking garage entrances or new areas where people are staying, and marking and deleting invalid nodes and edges when previously passable roads are blocked by temporary barriers. Fifth, branch weight adjustment is used, increasing the exploration weight of high-priority branches such as teaching buildings and dormitories, and decreasing the exploration weight of low-relevance branches such as logistics areas and parking lot interiors, ultimately resulting in a compact, efficient, and self-optimizing topology map suitable for campus search tasks.

[0051] 6) Based on the path-growing semantic topology map after structural self-optimization, a hierarchical navigation strategy is adopted to make decisions, output navigation control commands, and make judgments based on the execution status to obtain navigation execution feedback; specifically, hierarchical navigation decisions are executed based on the self-optimized topology map, and navigation control commands such as target nodes and robot speed and steering are output, while the command execution status and search results are collected simultaneously to generate navigation execution feedback.

[0052] Navigation execution feedback is based on a path-growing semantic topology map after structural self-optimization. Specifically, it uses a three-tiered navigation strategy—high-level frontier exploration node selection, mid-level path planning, and low-level dynamic obstacle avoidance—to make decisions based on the self-optimized topology map, outputting navigation control commands that include topology node selection and speed control. Navigation execution feedback is generated by collecting the execution status and results of the navigation control commands, including planning and obstacle avoidance results, frontier node selection status and call frequency, environmental change data, and motion execution deviation information. In this scenario, the specific implementation of the three-tiered navigation strategy is as follows: the high-level frontier exploration node selection unit extracts the semantic-spatial-task fusion features of the topology map, combined with branch exploration weights... The optimal frontier exploration node within the campus is selected as the long-field sub-target, and the large-scale campus search is broken down into sequential sub-target approximation tasks. The mid-level path planning unit plans the optimal travel path from the current node to the frontier sub-target based on the node connectivity and edge travel cost of the topological map. The bottom-level dynamic obstacle avoidance unit outputs robot speed and steering control commands based on real-time campus perception data to avoid dynamic pedestrians, non-motorized vehicles, and temporary obstacles within the campus. Simultaneously, the execution status of navigation control commands, obstacle avoidance results, environmental changes, and target detection status are collected to generate navigation execution feedback. The feedback includes planning and obstacle avoidance results, frontier node selection status and call frequency, campus environment change data, robot motion execution deviation, and target personnel detection information.

[0053] 7) Based on navigation execution feedback, determine the task completion status. If the target is not reached, retry; if the target is reached, the task ends, forming a closed-loop iteration that drives the dynamic correction of the topology map structure. Specifically, based on navigation execution feedback, determine whether the target personnel have been found. If the target is not found, iterate through the entire process to continuously advance the search. Once the target is found and located, the task is completed, forming a closed-loop iteration throughout the process. Simultaneously, drive the dynamic correction of the topology map based on information such as road accessibility.

[0054] Navigation execution feedback drives dynamic correction of the topology map structure. Specifically, it feeds back the planning and obstacle avoidance results from the navigation execution feedback to the topology self-optimization stage in real time. This updates the feature encoding of topology nodes and edges, marks invalid connected edges, adds cross-branch short-circuit paths, adjusts the dynamic obstacle risk of nodes and the passage cost of edges, triggering dynamic correction of the topology map. It also synchronously feeds back the selection status and call frequency of leading nodes, updating the task characteristics and state characteristics of nodes to achieve dynamic matching between the topology map and navigation decision requirements. Simultaneously, it judges the task execution status based on the navigation execution feedback; if the navigation target is not reached, the entire navigation process iterates; if the target is reached, the task ends, achieving a closed-loop operation of bidirectional coupling between topology and decision-making. In this scenario, navigation execution feedback drives… The specific implementation of dynamic topology correction is as follows: the planning and obstacle avoidance results are fed back in real time, the feature codes of campus topology nodes and edges are updated, invalid connected edges blocked by fences are marked, passable cross-branch short-circuit paths are added, the dynamic obstacle risk of nodes in densely populated areas and the passage cost of corresponding edges are adjusted, and dynamic correction of the topology map is triggered; the selection status and call frequency of leading nodes are fed back synchronously, the task characteristics and status characteristics of corresponding nodes are updated, and the dynamic matching of the topology map and campus search decision requirements is realized; at the same time, the navigation execution feedback is used to determine whether the target personnel have been found. If the target is not found, the entire process is iterated to continuously advance the campus area search. If the target personnel are found and located, the task ends. The entire process realizes the closed-loop operation of topology-decision bidirectional coupling.

[0055] This invention decouples the method from the robot hardware platform and multi-source sensing device types, making it adaptable to various mobile robots, including wheeled and tracked types. It eliminates the need for pre-building a global dense map, relying instead on robot motion state information and environmental perception information to complete online path-growing semantic topology map construction, node and edge information annotation and feature encoding, generation of a self-optimized path-growing semantic topology map, hierarchical navigation strategy decision-making, and navigation execution feedback loop. This effectively adapts to various indoor and outdoor unknown environments, meeting the long-field-of-view navigation requirements in complex and unknown environments. In this scenario, the... In addition to being compatible with wheeled campus security patrol robots, the method can also be adapted to tracked mobile robots for search tasks in mountainous areas and stairwells on campus. The entire process requires no pre-built dense map of the entire campus; relying on collected motion state information and environmental perception information, it can complete the entire closed loop of topological map construction, node and edge feature encoding, topological self-optimization, hierarchical navigation decision-making, and navigation execution feedback online. It is effectively adaptable to various unknown environments on campus, including indoor teaching buildings, libraries, dormitory corridors, and outdoor roads, playgrounds, and green areas, fully meeting the needs of campus security for large-scale, long-range personnel search and navigation.

[0056] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent substitutions, improvements, or variations made within the spirit and principles of the present invention should fall within the scope of protection of the present invention.

Claims

1. A path-growing semantic topology self-optimizing navigation decision-making closed-loop method driven by a large model in unknown environments, characterized in that... include: The robot's motion state information and environmental perception information are collected through multi-source sensing devices; Based on the robot's motion state information, the already traveled path is solved as the skeleton, and nodes and edges are incrementally generated to obtain a path-growing semantic topology map. Multimodal large models are used to encode, infer, and complete environmental perception information in multiple dimensions, and to obtain the node and edge information annotation and feature encoding of path-growing semantic topology map. Based on the annotation and feature encoding of node and edge information of the path-growing semantic topology map, spatial connectivity and semantic relevance are judged, and structural self-optimization is performed on the path-growing semantic topology map to obtain the structurally self-optimized path-growing semantic topology map. Based on the path-growing semantic topology map after structural self-optimization, a hierarchical navigation strategy is adopted to make decisions, output navigation control commands, and obtain navigation execution feedback based on the execution status. Based on navigation execution feedback, the task completion status is determined. If the goal is not reached, the task is retried; if the goal is reached, the task ends, forming a closed-loop iteration that drives the dynamic correction of the topology map structure.

2. The method according to claim 1, characterized in that, The multi-source sensing device acquires robot motion state information and environmental perception information. Specifically, the multi-source sensing device is compatible with LiDAR, visual sensors, IMU, and odometry, eliminating the need for pre-built global dense maps. The acquired data supports online topology construction, optimization, and navigation decisions throughout the entire process. The robot motion state information is used to solve the robot's already traveled path, providing a foundation for the skeleton construction of a path-growing semantic topology map. The environmental perception information is used for differentiated multi-dimensional encoding of multimodal large models, providing data support for the information annotation, feature encoding, and topology self-optimization of nodes and edges in the path-growing semantic topology map. The robot motion state information and environmental perception information are subsequently used for the construction of the path-growing semantic topology map.

3. The method according to claim 1 or 2, characterized in that, The path-growing semantic topology map is constructed by incrementally generating topology map nodes and edges, using the robot's already traveled path as the unique skeleton. During the construction process, only path nodes in the already traveled area, semantic nodes related to the surrounding tasks, and frontier exploration nodes are included, and irrelevant environmental information unrelated to the already traveled path is not stored.

4. The method according to claim 3, characterized in that, The path-growing semantic topology map node and edge information annotation and feature encoding are implemented through a multimodal large model based on environmental perception information. Specifically, the multimodal large model performs heterogeneous data parsing, reasoning, and prior completion on the environmental perception information, outputs semantic analysis conclusions, and completes the information annotation and differentiated multi-dimensional feature encoding of nodes and edges accordingly. During this process, the large model simultaneously completes the unified parsing of heterogeneous semantic data, fuzzy information reasoning, local semantic completion, task relevance scoring, and feature encoding verification and completion.

5. The method according to claim 4, characterized in that, The self-optimized path-growing semantic topology map is obtained based on the path-growing semantic topology map and its node and edge information annotation and feature encoding. Specifically, by combining the feature encoding information of nodes and edges, spatial connectivity and semantic relevance are judged, and the multimodal large model performs a real-time self-optimization operation to realize the dynamic optimization of the initially established path-growing semantic topology map, thus obtaining the self-optimized path-growing semantic topology map.

6. The method according to claim 1 or 5, characterized in that, The structural self-optimization operations specifically include: topology node integration, intelligent pruning, branch merging, cross-branch shorting, dynamic node addition and deletion, and branch weight adjustment. Among them, intelligent pruning removes invalid and redundant branches, branch merging simplifies similar nodes and adjacent edges, cross-branch shorting connects non-adjacent nodes through reasoning, dynamic node addition and deletion adapts to environmental changes, and branch weight adjustment allocates exploration priority according to task value. Finally, a compact, efficient, and adaptable self-optimized topology map that fits the current environment and navigation task is obtained.

7. The method according to claim 6, characterized in that, The navigation execution feedback is based on a path-growing semantic topology map after structural self-optimization. Specifically, it uses a three-tiered navigation strategy—high-level frontier exploration node selection, mid-level path planning, and low-level dynamic obstacle avoidance—to make decisions and output navigation control commands that include topology node selection and speed control. By collecting the execution status and results of the navigation control commands, navigation execution feedback is generated. The feedback includes planning and obstacle avoidance results, frontier node selection status and call frequency, environmental change data, and motion execution deviation information.

8. The method according to claim 7, characterized in that, The navigation execution feedback is used to drive the dynamic correction of the topology map structure. Specifically, the planning and obstacle avoidance results in the navigation execution feedback are fed back to the topology self-optimization stage in real time. This is used to update the feature encoding of topology nodes and edges, mark invalid connected edges, add cross-branch short-circuit paths, adjust the dynamic obstacle risk of nodes and the passage cost of edges, and trigger the dynamic correction of the topology map. The selection status and call frequency of the leading edge nodes are fed back synchronously to update the task characteristics and state characteristics of the nodes, so as to achieve dynamic matching between the topology map and the navigation decision requirements. At the same time, the task execution status is judged according to the navigation execution feedback. If the navigation target is not reached, the entire navigation process is iterated in a loop. If the target is reached, the task ends, so as to achieve a closed-loop operation of bidirectional coupling between topology and decision.

9. The method according to claim 1, characterized in that, The method described is decoupled from robot hardware platforms and multi-source sensing device types, and can be adapted to various mobile robots, including wheeled and tracked robots. It does not require the pre-construction of a global dense map. Relying on robot motion state information and environmental perception information, it completes the construction of a path-growing semantic topology map, the annotation and feature encoding of path-growing semantic topology map nodes and edges, the generation of a path-growing semantic topology map after structural self-optimization, the hierarchical navigation strategy decision-making, and the navigation execution feedback loop online. It effectively adapts to unknown environments in multiple indoor and outdoor scenarios and meets the long-field-of-view navigation requirements in complex and unknown environments.

Images