A jig body intelligent robot autonomous inspection and real-time situation awareness method

Abstract

This invention relates to the field of robot control technology, specifically to a method for autonomous inspection and real-time situational awareness of an embodied intelligent robot. The technical solution includes: adopting a three-layer collaborative architecture of edge-cloud. First, feasible regions on the map are calibrated. A multi-dimensional waypoint cost matrix is ​​generated through pre-calculation. A Pareto route solution set is obtained using a third-generation non-dominated sorting genetic algorithm. The optimal inspection route is selected by entropy weight approximation of the ideal solution. A three-level nested task model is constructed, enabling autonomous execution in weak network conditions via edge-cloud. A three-level hidden danger closed-loop system and a panoramic command screen are built, along with dual-mode pointer table calibration, a complete intelligent training iteration link, and a hash chain anti-tampering multi-tenant permission system. This solution balances full inspection coverage, time and energy consumption, adapts to dynamic obstacle scenarios, allows for continuous optimization of the detection model, and meets the needs of industrial inspection control and safety compliance.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a method for autonomous inspection and real-time situational awareness of an embodied intelligent robot. Background Technology

[0002] With the increasing demand for unmanned operations in industrial settings, quadruped inspection robots, with their strong terrain adaptability and flexible load configuration, have gradually replaced manual labor in high-risk and repetitive daily inspection tasks. Currently, most mainstream inspection scheduling platforms use a manual pre-programmed static route model, combined with a single vision detection algorithm for anomaly identification, but this approach still has core shortcomings in practical industrial applications.

[0003] The existing platform inspection routes only support simple waypoint settings, which cannot achieve fine configuration of road segment-level navigation parameters, nor can it embed continuous along-line detection in the travel path. Furthermore, the route planning logic is not designed in combination with the gait characteristics and terrain adaptability of quadruped robots. The inspection action and navigation process have low coupling, making it difficult to adapt to the differentiated needs of complex industrial scenarios. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for autonomous inspection and real-time situational awareness of an embodied intelligent robot, comprising the following steps: S1. Import the occupied grid map and 3D point cloud data generated by the embodied inspection robot on-site mapping, draw movable areas to generate feasible domain masks, and label terrain types to generate terrain cost matrices; establish a dual-state activation verification mechanism of expected activation and actual activation to ensure that the map configured in the cloud and the map actually loaded on the robot are synchronized and effective.

[0005] S2. The inspection route is abstracted into a weighted directed topology graph, with the inspection waypoints carrying detection actions as nodes and the travel paths with gait parameters and detection configurations as directed edges. The Pareto optimal route solution set is obtained by pre-calculating the path cost of all inspection waypoint pairs through the path planning algorithm, and finally outputting the comprehensive optimal inspection route.

[0006] S3. Based on the comprehensive optimal inspection route, construct a nested task model, configure scheduling rules and failure retry strategies; for multi-robot multi-task scenarios, use a multi-objective genetic algorithm with time window constraints to achieve optimal task allocation and execution sorting.

[0007] S4. All task data is sent to the site edge service unit. When the network is normal, the cloud and the edge are scheduled synchronously. When the network is interrupted, the edge autonomously executes the task according to the pre-stored rules and caches the data locally. After the network is restored, the execution status is synchronized through the breakpoint resume mechanism.

[0008] S5. Construct a three-level progressive data system, collect hidden dangers by equipment tag number through the equipment tree, and complete the closed-loop management of hidden dangers by adopting a three-state review process; and realize full-scene situation visualization with a panoramic command screen.

[0009] S6 features a built-in closed-loop machine learning operation and maintenance system that includes dataset management, online annotation, model training, accuracy evaluation, deployment, and performance sampling. For pointer-type instruments, it employs a linear + segmented dual-mode point calibration method to achieve accurate readings of non-standard dials, and continuously optimizes the accuracy of the detection model based on on-site inspection data.

[0010] Preferred: S2 includes: (1) topology inspection route construction; (2) adopting a multi-dimensional path pre-computation + sequence optimization decoupling architecture; (3) dynamic scene local route replanning.

[0011] Preferred method for constructing topology-based inspection routes includes: defining the inspection route as a weighted directed topology graph R=(V, P), and the set of waypoints V={v1,v2,...,v...} i ...v I ), where i is the number of each waypoint on the inspection route, and I is the total number of waypoints on the inspection route; the attribute vector of each waypoint is: v i ={x i ,y i ,yaw i ,t stayi G i ,T i priority i}, where (x i ,y i ) represents the world coordinates on the map, yaw i For the fuselage facing angle, t stayi G is the duration of stay at waypoints. i For the set of preset positions of the gimbal, T i For the set of mounted detection items, priority i For detection priority; the set of directed edges P = {p ii’ |v i ,v i’ V}, i and i' are waypoint numbers on the inspection route, p ii’ For each directed edge, the path between two waypoints is represented by the attribute vector: p ii’ ={d ii’ ,g ii’ ,v ii’ ,m ii’ ,dir ii’ ,S ii’ cost ii’}, where d ii’g represents the actual path length. ii’ For a quadrupedal walking stance, v ii’ Let m be the speed of travel. ii’ For path mode, dir ii’ S represents the direction of travel. ii’ For a continuous set of detections along the route, cost ii’ The overall cost of the route is calculated; a four-level validity check is performed after the route is laid out.

[0012] Preferred method: After the route is laid out, a four-level legality check is performed, including: 1. Connectivity check: The topology map is traversed using a depth-first search algorithm to confirm that all inspection waypoints are reachable from the starting point and that there are no isolated waypoints or broken paths; 2. Feasibility domain check: The feasible domain mask values ​​of all waypoint coordinates and path coverage grids are checked point by point. If all values ​​are 1, the check passes; 3. Detection integrity check: The checklist of mandatory inspection equipment is compared to confirm that all mandatory inspection equipment has the corresponding inspection items attached and there are no omissions; 4. Robot motion school check: The path slope and turning angle are checked to see if they are within the maximum motion capability threshold of the quadrupedal inspection robot. If they exceed the threshold, an adjustment prompt is given.

[0013] The preferred approach is to adopt a decoupled architecture of multi-dimensional path pre-computation and sequence optimization, which includes: the first stage: multi-dimensional path cost matrix generation, where path search and attribute calculation are performed independently for each waypoint pair, and finally an I×I multi-dimensional path cost matrix is ​​generated; the second stage: multi-objective sequence optimization, where the optimal waypoint visit sequence is solved with three minimization objectives as the optimization direction.

[0014] Preferred steps for generating the multi-dimensional path cost matrix include: 1. Constructing a complete optimal path raster sequence. ii’ 2. Calculation of multiple attributes along a path; Preferred: Constructing a complete and optimal path raster sequence. ii’ Includes: waypoint v i The corresponding grid cell serves as the starting point, and waypoint v i’ The corresponding grid cell is used as the endpoint. An A-star search is performed on the occupied grid cell map. The node evaluation function is: f(j) = g(j) + h(j), where j is the number of each grid cell on the path, and h(j) is the distance from the current grid cell to the target waypoint v. i’ The Euclidean distance between the coordinates; g(j) is the cumulative travel cost from the starting point to the current grid, which is updated gradually as the node expands. , where g(j prev ) represents the cumulative cost of the previous adjacent raster, r is the raster map resolution, and c(j) is the terrain cost coefficient of the current raster; the search terminates when it extends to the endpoint raster, and the parent node is traced back from the endpoint to obtain the result from v. i to v i’ Complete optimal path raster sequenceii’ .

[0015] Preferred: Multi-dimensional path and multi-attribute calculation, based on path raster sequence. ii’ Calculate the four core attributes of this path segment and store them all in the corresponding positions of the path cost matrix: actual path length d ii’ Average terrain cost coefficient c of road segment t (ii') Estimated travel energy consumption e ii’ Matrix filling encapsulates the above four attributes into a structure and stores it in the i-th row and i'-th column element CostMatrix(ii') of the path cost matrix; after all waypoint pairs are calculated, the complete I×I multi-dimensional path cost matrix is ​​obtained.

[0016] Preferred: Estimated travel energy consumption Where μ is the coefficient of ground friction, m is the mass of the robot, and g is the acceleration due to gravity (taken as 9.8 m / s²). 2 ), k g This is the motion loss coefficient (calibrated according to the corresponding model).

[0017] Preferred approach: Solving for the optimal waypoint visit sequence with three minimization objectives as the optimization direction: Objective 1: Minimize the total inspection time .

[0018] Preferred objective 2: Minimize total energy consumption Among them, e ii’ To estimate travel energy consumption; P idle This refers to the robot's standby power.

[0019] Preferred Target 3: Minimize the coverage rate of high-priority detections. , where N is the total number of high-priority mandatory checkpoints, and N' is the number of high-priority points covered by the route.

[0020] Preferred: The penalty function method is used to handle constraints, and a penalty term is added to the objective function value for solutions that violate the constraints. Where λ is the penalty coefficient, m is the constraint number, and M is the total number of constraints. The degree of violation of the m-th constraint.

[0021] Preferred approach: Employing the entropy weight-approximation ideal solution ranking method for multi-attribute decision-making, outputting the comprehensive optimal route. Specific steps: Step 1: Data standardization. Perform range standardization on the three target indicators to eliminate dimensional differences; the standardized numerical range is [0,1]. Step 2: Calculate indicator weights using the entropy weight method. Step 3: Determine the positive and negative ideal solutions; the positive ideal solution is a virtual solution where all indicators reach their optimal values, and the negative ideal solution is a virtual solution where all indicators reach their worst values. Step 4: Calculate the deviation distance of each solution from the positive and negative ideal solutions; the distance D to the positive ideal solution... k + Distance D to the negative ideal solution k - Step 5: Calculate and sort the relative proximity scores. Calculate the relative proximity score Ck for each scheme, using the formula: The proximity score ranges from [0,1], with a larger value indicating better overall performance. All schemes are sorted from largest to smallest by Ck, and the scheme with the highest proximity score is selected as the optimal inspection route.

[0022] Preferred: Local route replanning in dynamic scenes includes: 1. Obstacle detection and replanning triggering; 2. Delineating the local replanning window; 3. Using the improved D-Star Lite algorithm to search for the optimal path within the local window; 4. Path motion verification; 5. Path attribute update and power verification; 6. Task status synchronization.

[0023] Preferred: Step S3 includes: (1) a three-level nested task model; (2) an exponential backoff failure retry strategy; and (3) a multi-robot dynamic scheduling method.

[0024] The technical effects and advantages of this invention are as follows: 1. Significantly improved inspection reliability: The three-layer architecture of end-edge-cloud has a clear division of labor, with unified control in the cloud, autonomous execution at the edge, and precise execution at the end. It can still ensure the continuity of inspection operations even in network fluctuation scenarios, solving the industry pain point of inspection interruption in weak network environments, and making the system more adaptable to different environments.

[0025] 2. Improved scientificity and efficiency of route planning: The topology-based route model combines waypoint detection with continuous detection along the path, and deeply couples navigation gait parameters with detection actions; the two-stage decoupled multi-objective optimal planning algorithm, combined with the characteristics of quadruped robots, takes into account multiple dimensions of optimization such as time, energy consumption, and coverage, and greatly improves wiring efficiency and route economy.

[0026] 3. Enhanced task scheduling and collaboration: The three-level nested task orchestration model enables full-link control from macro scheduling to atomic actions. The multi-robot dynamic scheduling algorithm achieves optimal task allocation and load balancing, solving the scheduling conflict problem in multi-robot collaborative inspection, and improving overall inspection efficiency by more than 30%.

[0027] 4. Continuous Iterative Optimization of Detection Accuracy: Built-in complete machine learning operation and maintenance closed loop, the inspection data collected on site can be quickly transformed into model optimization samples. Combined with the dual-mode pointer meter calibration method, it can be adapted to various non-standard industrial dials, greatly improving the adaptability accuracy of artificial intelligence detection models to personalized on-site scenarios.

[0028] 5. Situational awareness and closed-loop control: The three-level data system combined with the panoramic command screen enables comprehensive perception of the inspection situation from macro to micro levels; the three-state hidden danger review process enables closed-loop handling of abnormal issues, improving operation and maintenance efficiency and the timeliness of problem handling.

[0029] 6. Sufficient security and compliance: The combination of four-level roles and attribute-based access control with multi-tenant isolation meets the permission management needs of multiple roles and customers in industrial scenarios; the hash chain anti-tampering audit log provides a reliable basis for compliance traceability and complies with industrial system security specifications.

[0030] 7. Multiple safety protections: The design includes feasible domain space constraints, task mutual exclusion mechanism, and remote triple parking protection, which avoid the safety risks of robot inspection and remote control from multiple levels of space, command and operation, and ensure the safety of on-site equipment and personnel. Attached Figure Description

[0031] Figure 1 This is a flowchart illustrating a method for autonomous inspection and real-time situational awareness of an embodied intelligent robot proposed in this invention.

[0032] Figure 2 This is a flowchart illustrating S2 in the autonomous inspection and real-time situational awareness method for an embodied intelligent robot proposed in this invention.

[0033] Figure 3 This is a flowchart illustrating the optimal route acquisition method in the autonomous inspection and real-time situational awareness method for an embodied intelligent robot proposed in this invention.

[0034] Figure 4 This is a flowchart illustrating the local route replanning process in a dynamic scene within an embodied intelligent robot autonomous inspection and real-time situational awareness method proposed in this invention.

[0035] Figure 5 This is a flowchart illustrating S3 in the autonomous inspection and real-time situational awareness method for an embodied intelligent robot proposed in this invention. Detailed Implementation

[0036] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the invention, and should not be construed as limiting the invention. Rather, embodiments of the invention include all variations, modifications, and equivalents falling within the spirit and scope of the appended claims.

[0037] Example 1 refer to Figure 1 This embodiment proposes a method for autonomous inspection and real-time situational awareness of embodied intelligent robots, implemented based on a three-layer collaborative architecture of edge-cloud: The end-side is a quadrupedal inspection robot that integrates a motion control unit, a visible light / thermal imaging multi-view camera, a controllable gimbal, and various types of sensor units, and is responsible for performing navigation, image acquisition, local reasoning, and status reporting. The edge is the site edge service unit, which is responsible for local robot scheduling, video stream transcoding, edge inference and on-site data caching, enabling autonomous execution of inspection tasks in weak network environments; The cloud platform is a cloud-based scheduling and management platform responsible for the control of all site equipment, map and route management, inspection task scheduling, full lifecycle management of detection models, data aggregation and analysis, situation visualization, and full-link permission governance.

[0038] The method specifically includes six core steps: S1. Map Construction and Feasible Domain Labeling: Import the occupied grid map and 3D point cloud data generated by the embodied inspection robot on-site mapping, draw movable areas to generate feasible domain masks, and label terrain types to generate terrain cost matrices; establish a dual-state activation verification mechanism of expected activation and actual activation to ensure that the map configured in the cloud and the map actually loaded on the robot are synchronized and effective. The feasible domain mask is a two-dimensional matrix with the same size as the grid map. A matrix element value of 1 represents that the corresponding location belongs to a movable area that the robot is allowed to pass through, and 0 represents a prohibited area; the terrain cost matrix contains basic cost coefficients corresponding to different terrains, used for path cost calculation; in the dual-state activation verification mechanism, the expected activation state is the map state that should be effective recorded in the cloud, and the actual activation state is the map state that is actually loaded and running reported by the robot. The map is considered to be officially effective only when the two are consistent, and can be used for navigation and inspection tasks. For example, control the quadruped embodied inspection robot to complete the mapping of the entire substation, export the mapping compressed package containing the occupied grid base map, coordinate parameter files and 3D point clouds, and upload it to the map management module of the cloud scheduling management platform. After importing the map, the system automatically parses the raster data, classifying all rasters into three categories: passable rasters, obstacle rasters, and unknown rasters. Maintenance personnel use the polygon tool in the 2D map workbench to draw passable areas such as equipment channels and equipment zones. The system automatically generates a feasible region mask: a 2D matrix with the same size as the raster map, where elements within the movable area have a value of 1, and elements within the prohibited area have a value of 0. Special terrain features such as steps, slopes, and rugged ground are marked on the map, generating a terrain cost matrix: flat ground cost coefficient 1.0, slope 1.5, steps 2.0, and rugged ground 1.8. The configured map is then sent to the corresponding robot. After the robot loads the map, it reports its actual activation status to the cloud. The cloud compares the expected activation status with the actual activation status, determining that the map is officially effective and can be used for subsequent navigation and inspection tasks.

[0039] S2. Topology-based Inspection Route Construction and Optimal Planning: The inspection route is abstracted as a weighted directed topology graph, with inspection waypoints carrying detection actions as nodes and travel paths with gait parameters and detection configurations as directed edges. A path planning algorithm is used to pre-calculate the path costs of all inspection waypoint pairs to solve for the Pareto optimal route solution set, ultimately outputting the comprehensive optimal inspection route. (Reference) Figure 2 The specific content includes: (1) a topological inspection route construction method; the inspection route is defined as a weighted directed topological graph R=(V, P), and the inspection waypoint set V={v1,v2,...,v... i ...v I ), where i is the number of each waypoint on the inspection route, and I is the total number of waypoints on the inspection route; the attribute vector of each waypoint is: v i ={x i ,y i ,yaw i ,tstayi G i ,T i priority i}, where (x i ,y i ) represents world coordinates on the map (unit: meters), yaw i For the fuselage facing angle (range -180° to 180°), t stayi G is the time spent at waypoints (in seconds). i For the set of preset positions of the gimbal, T i For the set of mounted detection items, priority i Detection priority (levels 0-3, with 3 being the highest). The set of directed edges P = {p ii’ |v i ,v i’ V}, i and i' are waypoint numbers on the inspection route, p ii’ For each directed edge, the path between two waypoints is represented by the attribute vector: p ii’ ={d ii’ ,g ii’ ,v ii’ ,m ii’ ,dir ii’ ,S ii’ cost ii’}, where d ii’ g represents the actual path length (in meters). ii’ For quadrupedal walking stances (basic standard / staircase standard / flat ground agility / staircase agility), v ii’ The speed of travel (unit: meters per second), m ii’ For path mode (straight line mode / autonomous obstacle avoidance mode), dir ii’ S represents the direction of travel (forward / backward). ii’ For a continuous set of detections along the route, cost ii’The overall travel cost of the route is calculated. For example, the inspection route is abstracted as a weighted directed topology graph R=(V,P), and the specific construction process is as follows: Inspection waypoint deployment: According to the substation equipment ledger, the observation positions of 12 devices to be inspected, such as switchgear and transformers, are set as the highest priority mandatory waypoints. Eight transition waypoints are added at the turning points of the route, for a total of 20 inspection waypoints. Each waypoint is configured with map world coordinates, aircraft orientation angle, and dwell time; mandatory waypoints are equipped with corresponding detection items and gimbal preset positions, while transition waypoints are not equipped with detection items and have a dwell time of 0. Directed edge generation of the travel path: Directed edges are automatically generated by connecting adjacent waypoints according to the initial inspection order. Staircase sections are configured with stair gait + low speed gear (0.3m / s), flat passages are configured with flat ground agile gait + normal speed gear (0.5m / s), and main roads are configured with safety helmet detection along the route. One frame of image is captured every 5 meters for detection. After the route is laid out, a four-level validity check is performed: 1. Connectivity check: A depth-first search algorithm is used to traverse the topology map to confirm that all inspection waypoints are reachable from the starting point, and there are no isolated waypoints or broken paths; 2. Feasibility domain check: The feasible domain mask values ​​of all waypoint coordinates and path coverage grids are checked point by point. If all values ​​are 1, the check passes; 3. Detection integrity check: The checklist of mandatory inspection equipment is compared to confirm that all mandatory inspection equipment has the corresponding inspection items attached, and there are no omissions; 4. Robot kinematics check: The path slope and turning angle are checked to see if they are within the maximum motion capability threshold of the quadrupedal inspection robot. If they exceed the threshold, adjustments are prompted. For example, in the fourth-level validity check: the route is fully connected after depth-first search verification, all waypoints and paths are within the feasible domain, all 12 mandatory inspection equipment have inspection items attached, and the path slope and turning angle meet the robot kinematic constraints. After the check passes, the route draft is saved.

[0040] (2) The decoupled architecture of path pre-computation + sequence optimization is adopted to greatly improve the efficiency of multi-objective planning: First stage: multi-dimensional path cost matrix generation; for each waypoint pair, path search and attribute calculation are performed independently to finally generate an I×I multi-dimensional path cost matrix. The specific steps are as follows: 1. Generate a multi-dimensional path cost matrix for waypoint v i The corresponding grid cell serves as the starting point, and waypoint v i’ The corresponding grid cell is used as the endpoint. An A-star search is performed on the occupied grid cell map. The node evaluation function is: f(j) = g(j) + h(j), where j is the number of each grid cell on the path, and h(j) is the distance from the current grid cell to the target waypoint v. i’ The Euclidean distance between the coordinates of the node and the current grid. g(j) is the cumulative travel cost from the starting point to the current grid, which is updated progressively as the node expands. , where g(j prev ) represents the cumulative cost of the previous adjacent grid cell, r is the grid map resolution (the actual length corresponding to a single grid cell), and c(j) is the terrain cost coefficient of the current grid cell. The search terminates when it extends to the endpoint grid cell, and the parent node is traced back from the endpoint to obtain the result from v. ito v i’ Complete optimal path raster sequence ii’ 2. Path multi-attribute calculation, based on path raster sequence. ii’ Calculate the four core attributes of this path segment and store them all in the corresponding positions of the path cost matrix: actual path length d ii’ ; , where J is the path p ii’ The total number of rasters included, where r is the raster map resolution. Average terrain cost coefficient c for road segments. t (ii') is the average value of the terrain cost coefficients for all grid cells along the path. Estimated travel time t ii’ Calculated based on preset gait and speed levels for the area along the route; details will not be elaborated here. Estimated energy consumption e ii’ Based on the energy consumption model of the quadruped robot, the following calculations were performed: Where μ is the coefficient of ground friction, m is the mass of the robot, and g is the acceleration due to gravity (taken as 9.8 m / s²). 2 ), k g This represents the motion loss coefficient (calibrated according to the corresponding aircraft model). Matrix filling encapsulates the above four attributes into a structure and stores it in the element CostMatrix(ii') in the i-th row and i'-th column of the path cost matrix. After calculating all waypoint pairs, a complete I×I multi-dimensional path cost matrix is ​​obtained. Subsequent sequence optimization does not require repeated path search; all segment attributes can be obtained simply by looking up the table, improving the algorithm iteration efficiency by two orders of magnitude. For example, for 20 waypoints to be inspected, a two-stage algorithm is used to find the optimal inspection sequence: All directed combinations of the 20 waypoints are traversed group by group, resulting in 380 waypoint pairs. Each group independently performs path search and attribute calculation on Satellite A. Taking the waypoint from gate v1 to switch cabinet v2 as an example, the search yields a path containing 240 grids, with an actual path length d12 = 240 × 0.05 = 12 meters; the path is entirely flat, with an average terrain cost coefficient of 1.0; the road speed is 0.5 m / s, and the estimated travel time t12 = 12 ÷ 0.5 = 24 seconds; substituting these values ​​into the energy consumption formula, the estimated travel energy consumption is calculated as e12 = (0.6 × 40 × 9.8 + 2.1 × 0.5). 2 )×12×1.0≈2829 Joules. After all waypoint pairs are calculated, a 20×20 multi-dimensional path cost matrix is ​​generated, which can be directly looked up and called in subsequent optimizations.

[0041] Phase 2: Multi-objective sequence optimization. The optimal waypoint visit sequence can be solved by minimizing three objectives: Objective 1: Minimize the total inspection time. That is, the sum of travel time across all routes and stop time at all waypoints. Objective 2: Minimize total energy consumption. Among them, e ii’ To estimate travel energy consumption; P idleRobot standby power (unit: watts). Objective 3: Minimize the coverage of high-priority detection: Where N is the total number of high-priority mandatory inspection points, and N' is the number of high-priority points covered by the route. The optimal solution must satisfy F3=0 (full coverage). Constraints and penalty functions: Hard constraints include: feasible region constraints, remaining power constraints (total energy consumption ≤ remaining power - safety reserve power), detection time window constraints, and robot kinematic constraints. A penalty function method is used to handle constraints, and a penalty term is added to the objective function value for solutions that violate the constraints. Where λ is the penalty coefficient (much larger than the order of magnitude of the objective function value, taking a value more than 100 times the order of magnitude of the objective function value), m is the constraint condition number, and M is the total number of constraint conditions. The degree of violation of the m-th constraint is used to ensure that infeasible solutions are eliminated during the evolution process. Optimal solution selection: After iteratively outputting the Pareto optimal solution set, the entropy-weighted approximation of the ideal solution ranking method is used for multi-attribute decision-making to output the comprehensive optimal route. (Reference) Figure 3 The specific steps are as follows: First, data standardization. Perform range standardization on the three target indicators to eliminate dimensional differences. The standardized values ​​will range from [0,1]. Second, calculate the indicator weights using the entropy weight method; details are omitted here. Third, determine the positive and negative ideal solutions. The positive ideal solution is a virtual solution where all indicators reach their optimal values, and the negative ideal solution is a virtual solution where all indicators reach their worst values. Since larger standardized values ​​are better, therefore: the positive ideal solution Z... + ={z1 + ,…z k + …z K +},z k + ={z 1k’ + ,z 2k’ + ,…z kk’ + …z KK’ +}: Take the maximum value of each indicator; negative ideal solution Z - ={z1 - ,…z k - …z K -},z k - ={z 1k’ - ,z 2k’ - ,…z kk’ - …z KK’ -} where k is the index of the Pareto optimal route, K is the total number of Pareto optimal routes, k=1,2,…K; k' is the index of the optimization index, K' is the total number of optimization indices; here K' can have 3 values, k'=1,2,3, k'=1→total time index; k'=2→total energy consumption index; k'=3→high-priority detection coverage rate index, z kk’ Let z be the weighted standardized value of the k-th scheme and the k'-th indicator. k’ + Let K' be the positive ideal value corresponding to the k'-th indicator. Among all possible solutions, the weighted standardized maximum value of the k'-th indicator is the best (a larger standardized indicator is better). For example, assuming there are 3 Pareto route solutions K=3, the weighted standardized matrix of the three indicators is as follows: Table 1. Weighted Standardization Matrix Calculate positive and negative ideal solutions: Total duration index (k'=1): Maximum value 0.8, minimum value 0.4 → z1 + =0.8, z1 - =0.4; Total energy consumption index (k'=2): Maximum value 0.9, minimum value 0.3 → z2 + =0.9, z2 - =0.3; Coverage index (k'=3): Maximum value 0.8, minimum value 0.5 → z3 + =0.8, z3 - =0.5. The final result is: Z + ={0.8,0.9,0.8},Z - ={0.4,0.3,0.5} takes the minimum value of each index. Fourth step: Calculate the deviation distance of each scheme from the positive and negative ideal solutions; the spatial distance from each scheme to the positive and negative ideal solutions can be calculated separately using Euclidean distance. Distance to the positive ideal solution: Distance to the negative ideal solution: , where D k + The smaller the value, the closer the solution is to the optimal level; D k - The larger the value, the further the solution is from the worst-case scenario. Step 5: Calculate and sort the relative proximity scores. Calculate the relative proximity score Ck for each solution using the following formula: The proximity score ranges from [0,1], with a larger value indicating better overall performance. All routes are sorted by Ck from largest to smallest, and the route with the highest proximity score is selected as the optimal inspection route. This method outputs an optimal inspection route with six advantages: it eliminates subjective bias through entropy weighting, ensuring full coverage of high-risk equipment; it combines quadrupedal gait and terrain energy consumption models to calculate routes that closely match real-world operating conditions, allowing for direct implementation; it provides three alternative routes: comprehensive, duration, and energy consumption; its two-stage decoupling architecture significantly improves computational speed; and it only involves local replanning when encountering temporary obstacles, maintaining the global detection sequence while balancing static optimality and on-site dynamic adaptability, resulting in significantly reduced inspection time and energy consumption. The calculation method for the optimal inspection route may also include: calculating the relative proximity score of the data deviation coefficients of the route routes in each Pareto optimal solution set. Where i is the number of each waypoint on the inspection route, I is the total number of waypoints on the inspection route; k' is the sequence number of the optimization index, and K' is the total number of optimization indexes; z ik’ For the route segment k' from waypoint i (inclusive) to i+1 (exclusive), optimize the index value, z ik’ The optimized index standard value for the route segment numbered k from waypoint i (inclusive) to i+1 (exclusive) can be standardized based on a threshold, which will not be elaborated here. k’0 This is the standard deviation unit value of the optimization index numbered k'. The value can be obtained by dividing historical data into levels. Generally, it is obtained by dividing the difference between the maximum and minimum values ​​by 10. Of course, other value settings are not excluded. Setting this value method conforms to the logic of weakening the high-quality data and amplifying the low-quality data, making the calculation more accurate.

[0042] (3) Dynamic scene local route replanning; when the onboard sensors of the quadrupedal inspection robot detect new obstacles on the path, the obstacle status of the local grid map is updated, and the improved D-star path replanning algorithm is used to replan only the local path from the current waypoint to the next waypoint without changing the overall waypoint visit sequence; after the replanning is completed, the remaining power is automatically checked, and if the remaining power is insufficient to complete the entire inspection, an early return warning is triggered. Reference Figure 4The specific steps include: 1. Obstacle detection and replanning triggering: The quadrupedal inspection robot's onboard LiDAR / depth camera scans the environment in front in real time. After detecting a new obstacle, it maps the obstacle's coordinates and outline to the global occupied grid map; updates the local feasible region mask: marks the grid covered by the obstacle as prohibited (sets the value to 0), and synchronously updates the cost value of the corresponding grid in the terrain cost matrix; verifies the obstacle position: if the obstacle falls within the currently executed path segment (current waypoint → next waypoint), local replanning is triggered; if the obstacle is not on the current path, only the map is updated and the path is not adjusted. 2. Delineating the local replanning window: Taking the robot's current real-time position as the starting point and the next target waypoint as the ending point, a rectangular local replanning window is delineated, with the window range extended to both sides of the path with a 2-meter safety margin; replanning is only performed in the grid map within the window, reducing the search space to improve calculation speed, while ensuring that the endpoint is fixed to the original target waypoint and does not change the global waypoint access sequence. 3. An improved D-Star Lite algorithm is used to search for the optimal path within a local window: The algorithm searches backward from the target waypoint, maintaining the same path cost function as the global A-Star algorithm, accumulating grid length and terrain cost to ensure that the cost calculation rules for the new path are consistent with the global path; during the search, feasible region constraints are strictly verified, prohibiting the path from crossing restricted grids and movable area boundaries; after the search is completed, a new path grid sequence from the robot's current position to the next waypoint is obtained, replacing the obstacle-affected parts of the original path. 4. Path motion verification: The newly generated local path is validated for validity: the maximum slope and minimum turning radius of the path are calculated to ensure they are within the quadruped robot's motion capability threshold; if the requirements are not met, the replanning window boundary is expanded and the search is repeated until a valid path is obtained. 5. Path Attribute Update and Power Verification: Calculate the actual length, average terrain cost coefficient, estimated travel time, and estimated travel energy consumption of the new path segment, and update the attribute values ​​of the corresponding waypoint pairs in the path cost matrix; recalculate the total energy consumption of the remaining inspection journey (current remaining path energy consumption + return trip to charging station energy consumption), and compare it with the robot's current remaining power: if the remaining power ≥ remaining total energy consumption + safety reserve power (default 20% of total capacity), then the power is considered sufficient, and the new path is issued to the robot for execution; if the remaining power is insufficient, an early return warning is triggered, the current inspection task is terminated, the shortest path from the current location to the charging station is automatically planned, and the robot is controlled to return to charging first to ensure equipment safety. 6. Task Status Synchronization: Local path adjustments only modify the travel trajectory of a single path segment, without changing the order of subsequent waypoints or detection configurations; all detection tasks are executed as originally planned; path adjustment records are synchronously uploaded to the cloud scheduling platform and stored in the task execution log for subsequent traceability.

[0043] S3. Based on the comprehensive optimal inspection route, a nested task model is constructed, and scheduling rules and failure retry strategies are configured. For multi-robot, multi-task scenarios, a multi-objective genetic algorithm with time window constraints is used to achieve optimal task allocation and execution ordering, realizing full-link control of tasks from macro-scheduling to micro-actions, while also supporting multi-robot collaborative scheduling. (Reference) Figure 5 Specifically, it includes: (1) a three-level nested task model; constructing a three-level nested structure of "task attribute layer - subtask layer - execution step layer" to realize the fine arrangement of inspection tasks: Task attribute layer: stores global task information, including basic task information, scheduling rules, task priority, and failure retry strategy; for example, configure the task name "routine daily inspection of substation", the type is routine inspection, and associate the optimal inspection route; the triggering method is timed scheduling, and it is executed at 8:00 every working day; failure retry strategy: maximum retry 2 times, basic backoff time is 300 seconds, retry limit is 1800 seconds, and the waiting time is calculated according to the exponential backoff formula. Subtask layer: divided into detection subtasks corresponding to inspection waypoints and navigation subtasks corresponding to travel paths, supports custom subtask pre-dependencies, and is executed serially according to the route order by default; for example, 20 waypoints correspond to 20 detection subtasks, and 19 paths correspond to 19 navigation subtasks; it is executed serially according to the route order by default, and the custom main transformer detection subtask is a high priority, with the pre-dependency being the completion of the entry navigation subtask. Execution Step Layer: Subtasks are broken down into atomic action units that can be directly executed by the quadrupedal inspection robot, such as gimbal rotation, image acquisition, and inference calculation. For example, each navigation subtask is broken down into three atomic steps: starting navigation, moving, and reaching waypoint; each detection subtask is broken down into five atomic steps: gimbal rotation to preset position, camera image acquisition, intelligent inference, result reporting, and gimbal return to position. (2) Exponential backoff failure retry strategy: When a task fails, the exponential backoff algorithm is used to calculate the waiting time for the nth retry to avoid frequent retries causing system congestion. Where t0 is the base backoff time (in seconds), and the retry waiting time is set to an upper limit threshold t. max When t n >t max Take t maxTo avoid excessively long backoff time leading to task delays. (3) Multi-robot dynamic scheduling method; For scenarios with multiple robots and multiple inspection tasks at a single site, a multi-objective genetic algorithm with time window constraints is used to solve the optimal scheduling scheme: Optimization objectives: Minimum total completion time: Minimum latest completion time of all tasks (longest total execution time among all robots); Minimum task timeout rate: Minimum proportion of tasks completed beyond the time window; Highest robot load balance: Minimum difference in task load among robots (1 - standard deviation of total execution time of each robot / average of total execution time of each robot, the closer to 1, the more balanced the load). Constraints: Single robot serial execution constraint, power threshold constraint, task time window constraint, area operation mutual exclusion constraint (only one robot is allowed to work in the same area at the same time). The algorithm uses a two-layer integer encoding: the first layer is the robot number assigned to the task, and the second layer is the task execution order of each robot; through selection, crossover, and mutation operations, iterative optimization is performed to output the Pareto optimal scheduling table.

[0044] S4. All task data is sent to the site edge service unit. When the network is normal, the cloud and the edge are scheduled synchronously. When the network is interrupted, the edge autonomously executes the task according to the pre-stored rules and caches the data locally. After the network is restored, the execution status is synchronized through the breakpoint resume mechanism. This solves the problem of task interruption caused by network fluctuations in industrial sites. Full Task Deployment: After task orchestration is completed in the cloud, the complete task step sequence, route data, and detection configuration are deployed to the site edge service unit, with local cache backup on the edge. Synchronous Scheduling Mode: When the network is normal, the cloud and edge synchronize task status, and robot execution data is uploaded to the cloud in real time for unified cloud management. Autonomous Execution Mode: When the network is interrupted, the edge service independently controls the robot to execute tasks according to pre-stored scheduling rules. All detection data and execution logs are cached locally without cloud intervention. Resume Synchronization After the network is restored, the edge automatically synchronizes the execution data during the network outage to the cloud in batches. The cloud verifies and aligns the execution status, eliminating the need to re-execute tasks and achieving seamless transition. Autonomous Anomaly Handling: When a task fails during a network outage, the edge automatically reassigns it according to a preset retry strategy. If the maximum number of retries is exceeded, it is marked as a failure and awaits cloud intervention.

[0045] S5. Construct a three-tiered progressive data system. Hazards are categorized by device tag number using a device tree, and a three-state review process is employed to complete closed-loop hazard management. Homepage Dashboard Overview: The platform homepage summarizes four core indicators: online sites, online devices, daily inspection volume, and active alarms. It simultaneously displays inspection volume trends, alarm level distribution, recent alarm list, and ongoing task list, supporting automatic / manual refresh. Maintenance personnel can easily grasp the overall inspection status upon login. Three-tiered Progressive Data System: Original Inspection Record Layer: Stores all original inspection data, including readings, confidence levels, occurrence time, and corresponding device information, serving as the lowest-level basic data source; Alarm Event Layer: Alarm events are generated when inspection results match preset rules, displayed in high / medium / low-risk categories, and can be filtered by alarm level; Hazard Discovery Item Layer: Duplicate or related alarm events are aggregated into hazards, categorized hierarchically by device tag number using a device tree. Each device node indicates the number of hazards awaiting review, and filtering by review status and severity is supported. Three-State Hazard Review Process: A three-state flow mechanism of "Pending Review - Confirmed - False Alarm" is adopted. After reviewing the on-site evidence images corresponding to the hazard, maintenance personnel mark genuine anomalies as confirmed and proceed to the subsequent maintenance and handling process; algorithm-incorrect judgments are marked as false alarms and excluded. The review results are updated in real time with statistical data, forming a closed-loop anomaly management system. Panoramic Command Screen: Adopting a dark full-screen layout, it integrates seven types of panels: equipment telemetry, 3D point cloud map, alarm center, detection statistics, equipment power consumption, shift task plan, and inspection route information. It automatically refreshes every 15 seconds, adapting to scenarios where monitoring room screens are constantly monitored, achieving a comprehensive overview of the entire situation on a single screen.

[0046] S6, built-in dataset management, online annotation, model training, accuracy evaluation, deployment and online, and effect sampling of machine learning operation and maintenance closed loop; for pointer instruments, a linear + segmented dual-mode point calibration method is adopted to achieve accurate reading of non-standard dials, and the accuracy of the detection model is continuously optimized based on on-site inspection data. The detection model is continuously optimized for the on-site scenario, and the reading adaptation problem of non-standard instruments is solved at the same time. (1) Dual-mode calibration method for pointer instruments; for pointer instruments with different scale characteristics in industrial sites, two calibration modes are provided to ensure reading accuracy. The specific technology is existing and will not be elaborated here. After calibration, the test angle is entered through the angle preview function to verify the conversion accuracy. After verification, it is put into use for inspection. (2) Machine learning operation and maintenance closed-loop process; construct a complete closed loop of "data-annotation-training-evaluation-deployment-sampling" to realize rapid iteration of detection models: Dataset management: supports two import methods: local compressed package upload and online platform pull, automatically counts the number of categories and annotation boxes, and divides the training set and validation set according to the proportion; Online annotation: connects to online visual annotation tools, supports creating new annotation tasks, drawing boxes online, and directly importing annotation results as training datasets; Model training: select the target dataset and training round, submit the lightweight object detection algorithm training task, the task is executed serially in a queue, and displayed in real time. Training progress and accuracy metrics; Accuracy evaluation: The trained model is evaluated using an independent validation set, and core metrics such as average accuracy and category accuracy reports are output as the basis for model deployment; Deployment: The evaluated model is deployed to the inference service framework. If deployment fails, the system automatically rolls back to the original version to ensure uninterrupted production inference; One-click rollback to historical versions based on deployment history is supported; Performance sampling: Real-time inference is supported by uploading on-site images. Detection results are dynamically filtered through a confidence threshold slider, and the detection box effect is visualized to quickly verify the model's adaptability to the on-site scene. The multi-tenant fine-grained permission governance mechanism meets the requirements of multi-tenant isolation and compliance auditing in industrial scenarios, serving as a security module for the system. It features a four-level hierarchical role system: Viewer, Operator, Tenant Administrator, and Super Administrator roles, with permissions progressively increasing from top to bottom. Platform menus and operation buttons are automatically tailored according to roles, achieving layered control of basic permissions. An attribute-based access control mechanism, building upon role permissions, assigns key-value pair attribute tags to users, enabling fine-grained resource-level authorization control. This complements role permissions and adapts to the fine-grained authorization needs of complex scenarios. Hash Chain Anti-Tampering Audit: All sensitive operations (remote control, task modification, model deployment, permission changes, etc.) generate audit logs, which are linked together by a hash chain to ensure the records are immutable. The hash function adopts the national cryptographic SM3 algorithm; it supports one-click verification of hash chain integrity, recalculating the full chain hash and comparing it with the final record hash to verify whether the record has been tampered with, meeting industrial compliance traceability requirements.

[0047] Full Task Deployment: After task orchestration is completed in the cloud, the complete task step sequence, route data, and detection configuration are deployed to the site edge service unit, with local cache backup on the edge. Synchronous Scheduling Mode: When the network is normal, the cloud and edge synchronize task status, and robot execution data is uploaded to the cloud in real time for unified cloud management. Autonomous Execution Mode: When the network is interrupted, the edge service independently controls the robot to execute tasks according to pre-stored scheduling rules. All detection data and execution logs are cached locally without cloud intervention. Resume Synchronization After the network is restored, the edge automatically synchronizes the execution data during the network outage to the cloud in batches. The cloud verifies and aligns the execution status, eliminating the need to re-execute tasks and achieving seamless transition. Autonomous Anomaly Handling: When a task fails during a network outage, the edge automatically reassigns it according to a preset retry strategy. If the maximum number of retries is exceeded, it is marked as a failure and awaits cloud intervention.

[0048] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0049] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for autonomous inspection and real-time situational awareness of an embodied intelligent robot, characterized in that, The method for autonomous inspection and real-time situational awareness of the embodied intelligent robot includes: S1. Import the occupied grid map and 3D point cloud data generated by the on-site mapping of the embodied inspection robot, draw the movable area to generate a feasible region mask, and mark the terrain type to generate a terrain cost matrix; establish a dual-state activation verification mechanism of expected activation and actual activation to ensure that the map configured in the cloud and the map actually loaded on the robot are synchronized and effective. S2. The inspection route is abstracted into a weighted directed topology graph, with the inspection waypoints carrying detection actions as nodes and the travel paths with gait parameters and detection configurations as directed edges. The path cost of all inspection waypoint pairs is pre-calculated using a path planning algorithm to solve for the Pareto optimal route solution set, and finally the comprehensive optimal inspection route is output. S3. Based on the comprehensive optimal inspection route, construct a nested task model and configure scheduling rules and failure retry strategies; for multi-robot multi-task scenarios, use a multi-objective genetic algorithm with time window constraints to achieve optimal task allocation and execution ordering. S4. All task data is sent to the site edge service unit. When the network is normal, the cloud and the edge are scheduled synchronously. When the network is interrupted, the edge autonomously executes the task according to the pre-stored rules and caches the data locally. After the network is restored, the execution status is synchronized through the breakpoint resume mechanism. S5. Construct a three-level progressive data system, collect hidden dangers by equipment tag number through the equipment tree, and complete the closed-loop management of hidden dangers by adopting a three-state review process. S6 features a built-in closed-loop machine learning operation and maintenance system that includes dataset management, online annotation, model training, accuracy evaluation, deployment, and performance sampling. For pointer-type instruments, it employs a linear + segmented dual-mode point calibration method to achieve accurate readings of non-standard dials, and continuously optimizes the accuracy of the detection model based on on-site inspection data.

2. The method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 1, characterized in that, The S2 includes: (1) topology inspection route construction; (2) adopting a multi-dimensional path pre-computation + sequence optimization decoupling architecture; and (3) dynamic scene local route replanning.

3. The method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 2, characterized in that, The topology-based inspection route construction method includes: defining the inspection route as a weighted directed topological graph R=(V, P), and the inspection waypoint set V={v1,v2,...,v...} i ...v I ), where i is the number of each waypoint on the inspection route, and I is the total number of waypoints on the inspection route; the attribute vector of each waypoint is: v i ={x i ,y i ,yaw i ,t stayi G i ,T i priority i }, where (x i ,y i ) represents the world coordinates on the map, yaw i For the fuselage facing angle, t stayi G is the duration of stay at waypoints. i For the set of preset positions of the gimbal, T i For the set of mounted detection items, priority i For detection priority; the set of directed edges P = {p ii’ |v i ,v i’ V}, i and i' are waypoint numbers on the inspection route, p ii’ For each directed edge, the path between two waypoints is represented by the attribute vector p. ii’ ={d ii’ ,g ii’ ,v ii’ ,m ii’ ,dir ii’ ,S ii’ cost ii’ }, where d ii’ g represents the actual path length. ii’ For a quadrupedal walking stance, v ii’ Let m be the speed of travel. ii’ For path mode, dir ii’ S represents the direction of travel. ii’ For a continuous set of detections along the route, cost ii’ The overall cost of the route is calculated; a four-level validity check is performed after the route is laid out.

4. The method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 3, characterized in that, After the route is deployed, a four-level legality check is performed, including:

1. Connectivity check: The topology map is traversed using a depth-first search algorithm to confirm that all inspection waypoints are reachable from the starting point and that there are no isolated waypoints or broken paths; 2. Feasibility domain check: The feasible domain mask values ​​of all waypoint coordinates and path coverage grids are checked point by point. If all values ​​are 1, the check passes; 3. Detection integrity check: The checklist of mandatory inspection equipment is compared to confirm that all mandatory inspection equipment has the corresponding inspection items attached and there are no omissions; 4. Robot motion school check: The path slope and turning angle are checked to see if they are within the maximum motion capability threshold of the quadrupedal inspection robot. If they exceed the threshold, an adjustment prompt is given.

5. The method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 2, characterized in that, The decoupled architecture of path pre-computation + sequence optimization includes: the first stage: multi-dimensional path cost matrix generation, where path search and attribute calculation are performed independently for each waypoint pair, and finally an I×I multi-dimensional path cost matrix is ​​generated; the second stage: multi-objective sequence optimization, where the optimal waypoint visit sequence is solved with three minimization objectives as the optimization direction.

6. The method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 5, characterized in that, The specific steps for generating a multi-dimensional path cost matrix include:

1. Constructing a complete optimal path raster sequence. ii’ 2. Multi-dimensional path and multi-attribute calculation.

7. The method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 6, characterized in that, Constructing a complete optimal path raster sequence ii’ Includes: waypoint v i The corresponding grid cell serves as the starting point, and waypoint v i’ The corresponding grid cell is used as the endpoint. An A* search is performed on the grid cell map. The search terminates when it extends to the endpoint grid cell. The parent node is then traced back from the endpoint to obtain the result from v. i to v i’ Complete optimal path raster sequence ii’ .

8. The method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 6, characterized in that, Multi-dimensional path multi-attribute calculation, based on path raster sequence. ii’ Calculate the four core attributes of this path segment and store them all in the corresponding positions of the path cost matrix: actual path length d ii’ Average terrain cost coefficient c of road segment t (ii') Estimated travel energy consumption e ii’ Matrix filling encapsulates the above four attributes into a structure and stores it in the i-th row and i'-th column element CostMatrix(ii') of the path cost matrix; after all waypoint pairs are calculated, the complete I×I multi-dimensional path cost matrix is ​​obtained.

9. A method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 6, characterized in that, The steps for obtaining the optimal route include: Step 1: Data standardization: Perform range standardization on the three target indicators to eliminate dimensional differences; the standardized values ​​will range from [0,1]. Step 2: Calculate the indicator weights using the entropy weight method. Step 3: Determine the positive and negative ideal solutions. Step 4: Calculate the deviation distance of each solution from the positive and negative ideal solutions; obtain the distance D from the positive ideal solution. k + Distance D to the negative ideal solution k - Step 5: Calculate the relative closeness and rank the schemes. Calculate the relative closeness C of each scheme. k ;Assign all schemes to C k Sort the routes from largest to smallest, and select the route with the highest proximity as the overall optimal inspection route.

10. A method for autonomous inspection and real-time situational awareness of an embodied intelligent robot according to claim 2, characterized in that, The local route replanning in dynamic scenes includes:

1. Obstacle detection and replanning triggering; 2. Delineating the local replanning window; 3. Using the improved D-Star Lite algorithm to search for the optimal path within the local window; 4. Path motion verification; 5. Path attribute update and power verification; 6. Task status synchronization.

Images