Automatic path-finding intelligent robot and path-finding method thereof

By acquiring multidimensional spatial data to construct a topological map and utilizing a path planning model based on manifold learning, the problem of low efficiency of traditional algorithms in complex environments is solved, enabling robots to navigate efficiently and stably in dynamic environments.

CN121430620APending Publication Date: 2026-01-30CHENGXUN ELECTRONIC TECHNOLOGY (CHANGZHOU) CO LTD
View PDF 0 Cites 4 Cited by

Patent Information

Application Number
CN202511557070.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-10-29
Publication Date
2026-01-30

AI Technical Summary

Technical Problem

Traditional automatic pathfinding algorithms are inefficient in complex environments, struggle to find the globally optimal path, and are prone to getting stuck in local optima in dynamically changing environments, leading to unstable robot operation and low efficiency.

Method used

Multidimensional spatial data is acquired through environmental sensing devices to construct a hierarchical topological map. A path planning model based on manifold learning is used, combined with regional traffic weights and dynamic path optimization, to generate a smooth final execution path.

Benefits of technology

It improves the efficiency and safety of robot path planning in complex environments, enhances the robot's adaptability to dynamic obstacles, and ensures that the robot operates efficiently and stably in changing environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121430620A_ABST
    Figure CN121430620A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of robot navigation, and discloses an automatic path-finding intelligent robot and a path-finding method thereof. According to the method, multi-dimensional spatial data including obstacle distribution, ground features and dynamic target motion information is acquired through an environment sensing device, spatial discretization processing is performed on the data, and a topological map which has a hierarchical structure and includes a node connection relation and a region passing weight is generated. Then, according to the node connection relation of the topological map, a path planning model based on manifold learning is established, and an optimal path candidate set is determined by calculating the manifold distance between adjacent nodes; then extracting a key node sequence in the candidate set, dynamically optimizing a path in combination with a region passing weight, and generating a preliminary navigation track; and finally, inputting the initial track into a motion control model, carrying out smooth processing according to kinematics constraint of the robot, and outputting a final execution path. According to the invention, the robot can efficiently find the way in a complex environment.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of robot navigation technology, specifically to an intelligent robot with automatic pathfinding and its pathfinding method. Background Technology

[0002] The development of automated pathfinding technology is an ongoing process, its origins traceable to early, simple, pre-set path patterns. In the early stages of technological development, the application scenarios of robots were relatively limited, mostly in environments with simpler conditions and fewer changing factors, such as material handling in specific processes on early industrial production lines. These robots only needed to follow pre-set fixed routes to complete tasks, relying on simple programming instructions to move from the starting point to the target point. This method could meet basic production needs in the specific environment at that time.

[0003] With the advancement of technology and the expansion of application scenarios, people have placed higher demands on the autonomy and adaptability of robots, and automatic pathfinding technology has entered a new stage of development. In this stage, various algorithm-based pathfinding methods have emerged, among which map-building and search algorithm-based pathfinding is a typical example. Robots collect information about their surrounding environment through sensors, construct simple map models, and then use classic search algorithms such as Dijkstra's algorithm and A* algorithm to find a path from their current location to the target location on the map. These algorithms have improved the robot's pathfinding ability in relatively complex environments to a certain extent, enabling robots to cope with scenarios with simple obstacle distributions.

[0004] With the rapid development of technologies such as artificial intelligence and machine learning, automated pathfinding technology has ushered in an intelligent transformation. Modern automated pathfinding intelligent robots can not only perceive environmental information but also perform deep analysis and learning on this information, thereby achieving more intelligent and efficient path planning. For example, neural network models based on deep learning allow robots to learn pathfinding patterns in different scenarios from a large amount of environmental data, enhancing their adaptability in complex and ever-changing environments. This intelligent pathfinding technology lays a solid foundation for the widespread application of robots in more fields.

[0005] In traditional automatic pathfinding algorithms, exhaustive search can theoretically find all possible paths, but this approach suffers from serious efficiency problems in practical applications. This is because exhaustive search requires exploring every possible path in the solution space, and its time complexity typically grows exponentially with the problem size. In large-scale, complex environments, the computational load becomes extremely large, requiring the robot to expend significant time and computational resources to complete path planning, which is unacceptable in many scenarios with high real-time requirements.

[0006] While heuristic search algorithms, such as the A* algorithm, introduce heuristic functions to guide the search direction and improve search efficiency to some extent, they still have limitations in complex scenarios. The design of heuristic functions often relies on prior assumptions and simplified models of the environment. When the actual environment deviates significantly from these assumptions, the heuristic function may fail to accurately evaluate the value of nodes, causing the algorithm to get trapped in local optima. For example, in environments with irregularly distributed and dynamically changing obstacles, heuristic search algorithms may miss the globally optimal path due to inaccurate heuristic functions, or frequently perform invalid searches during the search process, reducing pathfinding efficiency. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent robot for automatic pathfinding and its pathfinding method, so as to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides an intelligent robot for automatic pathfinding and a pathfinding method thereof, the method comprising: The robot acquires multi-dimensional spatial data of its current environment through an environmental sensing device. The multi-dimensional spatial data includes obstacle distribution information, ground feature information, and dynamic target motion information. The multidimensional spatial data is spatially discretized to generate a topological map with a hierarchical structure, which includes node connection relationships and regional access weights. Based on the node connection relationships of the topology map, a path planning model based on manifold learning is established. The path planning model determines the optimal path candidate set by calculating the manifold distance between adjacent nodes. Extract the key node sequence from the optimal path candidate set, combine it with the regional traffic weight to perform dynamic path optimization, and generate a preliminary navigation trajectory. The initial navigation trajectory is input into the motion control model, which smooths the trajectory based on the robot's kinematic constraints and outputs the final execution path.

[0009] Preferably, the acquisition of multi-dimensional spatial data of the robot's current environment through the environmental sensing device includes: The environment is collected using lidar, and the three-dimensional point cloud data includes the geometric contours and height information of obstacles. Ground texture features and color markers for dynamic targets are extracted using a visual sensor. The ground texture features are used to distinguish between passable and inaccessible areas. By integrating the 3D point cloud data with ground texture features, a local environment coordinate system is constructed with the robot's current position as the origin. This local environment coordinate system is used to unify the data space benchmark.

[0010] Preferably, the step of spatially discretizing the multidimensional spatial data to generate a topological map with a hierarchical structure includes: The local environment coordinate system is divided into several grid cells of equal size, and each grid cell records the probability of obstacle occupancy and the ground friction coefficient. Based on the connectivity between grid cells, adjacent free grid cells are aggregated to form a super node, which represents the topological center of the passable area. Based on the Euclidean distance between the supernodes and the ground friction coefficient, the regional travel weight is calculated, and the regional travel weight is used to quantify the path selection cost.

[0011] Preferably, establishing a path planning model based on manifold learning includes: Select a starting node and a target node in the topology map. The starting node corresponds to the robot's current position, and the target node is input by the user or specified by the task planning system. The high-dimensional spatial relationships of supernodes are reduced using a diffusion mapping algorithm to generate a low-dimensional manifold embedding representation. In a low-dimensional manifold space, geodesic distance is used instead of Euclidean distance to redefine the connection cost between nodes, which is used to update the optimal path candidate set.

[0012] Preferably, extracting the key node sequence from the optimal path candidate set includes: Time series analysis is performed on the candidate set of optimal paths to screen node combinations that satisfy the motion continuity constraint; Based on the dynamic target motion information, predict the path conflict risk within the future time window and remove node sequences with a collision probability; The remaining node sequences are sorted in ascending order according to the regional access weight, and the first K sequences are selected as the input for the initial navigation trajectory.

[0013] Preferably, the dynamic path optimization based on the regional access weight includes: Real-time monitoring of the dynamic obstacle positions updated by the environmental sensing device, and adjustment of the passage weight of the affected grid cells; A rolling time-domain optimization strategy is adopted to recalculate the local paths of the key node sequence at fixed time intervals; When a sudden obstacle is detected, a path replanning mechanism is triggered, which generates an alternative trajectory based on the latest topology map.

[0014] Preferably, inputting the preliminary navigation trajectory into the motion control model includes: The curvature change points and velocity constraints of the preliminary navigation trajectory are analyzed to generate a piecewise polynomial fitting function; Transition points are inserted into the piecewise polynomial fitting function to ensure continuous robot acceleration; Based on the steering angle limitations of the wheel drive mechanism, the feasibility of the transition connection point is verified, and unexecutable trajectory segments are eliminated.

[0015] Preferably, the motion control model smooths the trajectory based on robot kinematic constraints, including: Establish the kinematic equations of the robot chassis, which include the maximum linear velocity, angular velocity, and acceleration / deceleration limits; The final execution path is discretized into dense waypoints, and the difference in motion parameters between adjacent waypoints is calculated. If the difference in motion parameters exceeds the threshold, transition waypoints are generated using Bézier curve interpolation until all waypoints satisfy the kinematic constraints.

[0016] Preferably, the method further includes: Multiple checkpoints are set on the final execution path, and these checkpoints are used to trigger local data updates from the environmental sensing device. When the robot reaches the checkpoint, it compares the difference between the actual environmental data and the topology map. If the difference exceeds the tolerance, it pauses execution and starts the map correction process.

[0017] Preferably, the map correction process includes: Reacquire lidar point cloud data and visual texture features at the current location to generate a corrected local environment coordinate system; The supernode connection relationships in the topology map are updated based on the correction results, and the passage weights of the affected areas are recalculated. Based on the updated topology map, the second half of the final execution path is regenerated starting from the current checkpoint.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention acquires multi-dimensional spatial data of the robot's current environment through an environmental sensing device, encompassing obstacle distribution information, ground feature information, and dynamic target motion information. This allows the robot to comprehensively and accurately understand its surroundings. In practical applications, such as logistics warehouses, the robot can use this multi-dimensional spatial data to clearly identify the location of static obstacles such as shelves and stacks of goods, and can also promptly detect uneven or damp ground conditions. It can also monitor the movement trajectories and speeds of other dynamic targets operating in the warehouse, such as robots and forklifts. Compared to traditional robot pathfinding methods that only perceive obstacles, this invention significantly reduces the risk of collisions between the robot and obstacles, improving its safety and reliability in complex environments. For example, in narrow warehouse aisles, the robot can plan its path in advance based on the perceived multi-dimensional spatial data, avoiding collisions with oncoming forklifts and ensuring the smooth transport of goods.

[0019] This invention spatially discretizes multidimensional spatial data to generate a hierarchical topological map and establishes a path planning model based on manifold learning. This approach enables the robot to quickly and efficiently find the optimal path candidate set. The hierarchical structure and node connections of the topological map provide the robot with a clear environmental framework, while the manifold learning-based path planning model filters the optimal path candidate set from numerous possible paths by calculating the manifold distance between adjacent nodes. Taking a hospital scenario as an example, the internal environment of a hospital is complex, with numerous departments and crisscrossing corridors. Using the pathfinding method of this invention, the robot can quickly plan the optimal path from the pharmacy to the ward based on the topological map and the path planning model, avoiding densely populated areas and other obstacles. Traditional pathfinding methods may require a lot of time for path searching and may even get stuck in local optima, failing to find the globally optimal path. This invention greatly improves the efficiency and quality of path planning, enabling the robot to reach the target location in a shorter time and improving overall work efficiency.

[0020] When generating the initial navigation trajectory, this invention extracts the key node sequence from the optimal path candidate set and performs dynamic path optimization by combining it with regional traffic weights. This allows the robot to flexibly adjust its path in complex and dynamic environments to adapt to constantly changing environmental conditions. In environments with frequent pedestrian traffic, such as shopping malls, the robot's originally planned path may be blocked by sudden crowds. In this case, the robot can adjust its path in a timely manner based on regional traffic weights, selecting areas with smoother traffic flow and avoiding congested areas. Through dynamic path optimization, the robot can respond to environmental changes in real time and maintain an efficient operating state. Compared with traditional fixed path planning methods, the robot of this invention has stronger environmental adaptability and can successfully complete tasks under various complex conditions, improving the robot's practicality and reliability in real-world applications.

[0021] The initial navigation trajectory is input into the motion control model, which smooths the trajectory based on the robot's kinematic constraints and outputs the final execution path. This process makes the robot's movement smoother, reduces wear on mechanical parts, and improves the robot's stability and reliability. On industrial production lines, robots need to move and operate frequently. If the movement is not smooth, it will not only affect production efficiency but may also lead to product quality problems. Using the motion control model of this invention, the robot can run along a smooth trajectory, reducing the impact on mechanical parts caused by sudden stops and starts. For example, when handling precision parts, smooth movement can prevent damage to the parts due to vibration, ensuring the smooth progress of the production process. Furthermore, smooth movement can also reduce noise during robot operation, creating a quieter working environment. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the working principle of the automatic pathfinding intelligent robot and its pathfinding method described in this invention. Figure 2 A flowchart of a method for acquiring multidimensional spatial data; Figure 3 A flowchart for discretization of multidimensional spatial data and generation of topological maps. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figure 1 This invention provides an intelligent robot for automatic pathfinding and a pathfinding method thereof, the method comprising: Multidimensional spatial data of the robot's current environment is acquired through an environmental sensing device. This data includes obstacle distribution information, ground feature information, and dynamic target motion information. The multidimensional spatial data collected by the environmental sensing device is spatially discretized to generate a hierarchical topology map, which includes node connectivity and region access weights. Based on the node connectivity in the topology map, a path planning model based on manifold learning is established. This model determines the optimal path candidate set by calculating the manifold distance between adjacent nodes. Key node sequences are extracted from the optimal path candidate set, and dynamic path optimization is performed using region access weights to generate a preliminary navigation trajectory. This preliminary navigation trajectory is input into a motion control model, which smooths the trajectory according to the robot's kinematic constraints and outputs the execution path.

[0025] Example 1: See Figure 2 Multidimensional spatial data includes obstacle distribution information, ground feature information, and dynamic target motion information. The environmental perception device integrates lidar and visual sensors to work together. The lidar collects three-dimensional point cloud data of the environment. The three-dimensional point cloud data contains the geometric contours and height information of obstacles. The lidar generates point clouds by emitting pulsed laser beams and measuring reflection time. Each point cloud data point contains three-dimensional coordinates and reflection intensity values. The point cloud data is filtered to remove noise and voxel grid downsampling is used to reduce the amount of data while retaining key features. Obstacle distribution information is extracted from the point cloud and independent obstacle entities are identified by clustering algorithms such as Euclidean clustering. The clustering algorithm groups nearby points based on the distance threshold between points, and each group represents an obstacle.

[0026] Obstacle geometry is calculated using the convex hull algorithm to determine boundary points. Height information is derived from the Z-coordinate of the point cloud. The vision sensor simultaneously captures RGB images and depth information. It extracts ground texture features and color markers for dynamic targets. Ground texture features are used to distinguish between passable and non-passable areas. Texture feature extraction employs a local binary mode algorithm to analyze image pixel neighborhood patterns. Color marker recognition is based on HSV color space segmentation. Dynamic targets are detected using a background subtraction algorithm. Motion vectors are estimated using inter-frame optical flow. Dynamic target motion information includes velocity direction and displacement. Trajectory is predicted using Kalman filtering. A local environment coordinate system with the robot's current position as the origin is constructed by fusing 3D point cloud data and ground texture features. This local environment coordinate system is used to unify the data space reference. The data fusion process involves sensor calibration. The extrinsic matrices of the LiDAR and vision sensor are obtained using a checkerboard calibration method. The intrinsic matrix corrects lens distortion. A coordinate transformation algorithm projects the point cloud data onto the image plane, and image features are mapped back to 3D space. A time synchronization mechanism uses hardware triggering to ensure data alignment. The establishment of the local environment coordinate system depends on the robot's localization system.

[0027] The positioning system provides real-time pose data. The origin of the coordinate system is set at the robot's center, with the X-axis along the forward direction, the Y-axis to the left, and the Z-axis vertically upward. Coordinate system transformation is achieved through a homogeneous transformation matrix. Multidimensional spatial data is stored in the form of point cloud maps and feature maps. The data acquisition process is continuous. The environmental perception device updates at a frequency of 10 frames per second. The LiDAR has a 360-degree scanning range, and the visual sensor has a 120-degree field of view. Obstacle distribution information is updated synchronously with the sensor. Ground feature information includes texture type and roughness. The visual sensor uses a convolutional neural network to classify ground materials. Dynamic target motion information tracks multiple targets. Data association algorithms solve target ID switching. The local environmental coordinate system is dynamically adjusted to adapt to robot motion. Extended Kalman filtering fuses wheel odometry and IMU data to improve coordinate system stability. Multidimensional spatial data is compressed and encoded to reduce storage overhead. Data stream pipeline buffering avoids data loss. The environmental perception device is calibrated periodically, and the calibration process is automated without interrupting the task. The timestamp alignment accuracy of the LiDAR and visual sensor reaches the millisecond level. The integrity and consistency of multidimensional spatial data are verified through checksums. The entire data acquisition system is embedded, with low latency and high throughput supporting real-time path planning.

[0028] The working details of the vision sensor are further elaborated. The vision sensor employs a global shutter camera to avoid motion blur, with an image resolution set at 1920x1080 pixels. Exposure time adapts to ambient lighting. Ground texture feature extraction utilizes… Filter banks enhance texture response; texture feature vectors are dimensionality-reduced and input into a support vector machine classifier, which outputs a passability probability value. Color markings for dynamic targets are predefined: red marks represent moving obstacles, and green marks represent safe zones. Color segmentation is thresholded in HSV space. Dynamic target tracking uses the Hungarian algorithm to associate detection boxes. Motion information prediction is based on a constant velocity model. Visual sensor data preprocessing includes white balance and gamma correction. Image pyramids are used to construct multi-scale features. Feature point matching uses ORB descriptors. When fusing 3D point cloud data with image data, the point cloud is projected onto the image to generate a depth map. The depth map is then compared with the RG... Image alignment, stereo vision triangulation to fill missing depth, accuracy assessment of local environmental coordinate system through reprojection error, coordinate system update frequency consistent with positioning system, continuous coordinate system transformation during robot movement, use of quaternions to represent rotation to avoid gimbal lock, historical data caching of multidimensional spatial data for playback analysis, thread-safe design of data acquisition module, support for asynchronous acquisition by multiple sensors, configurable LiDAR point cloud density, adjustable visual sensor frame rate, optimized power consumption of environmental perception device suitable for mobile platforms, occasional frame loss during sensor data time interpolation, strong robustness of the entire system, stable operation in changing lighting and dynamic environments.

[0029] The data fusion algorithm is implemented in the following ways: the coordinate transformation algorithm uses the least squares method to solve for the optimal transformation matrix; the point cloud data and image feature points are matched using the iterative nearest point algorithm for precise registration; the construction of the local environment coordinate system includes scale factor estimation to address scale differences between sensors; the fused data generates a unified point cloud; texture and color information is added to the point cloud; obstacle distribution information is visualized; ground feature information is rasterized; dynamic target motion information is smoothed over time; the local environment coordinate system is associated with the world coordinate system, and global reference is provided through GPS or SLAM; the data fusion accuracy reaches the centimeter level; real-time performance meets the requirements of the control loop; the environmental sensing device interface is standardized, supporting plug-in sensor expansion; the calibration algorithm runs online to dynamically correct sensor deviations; multi-dimensional spatial data compression uses an octree structure for efficient storage and retrieval; the data acquisition process is logged for easy fault diagnosis; the entire implementation method ensures data reliability and real-time performance, providing a solid foundation for subsequent processing.

[0030] A detailed description of LiDAR data acquisition is provided. The LiDAR model used is a 16-line rotating type with a scanning frequency of 10Hz and a point cloud density of 30,000 points per frame. The data format includes XYZ coordinates and reflectivity. Point cloud preprocessing includes outlier removal and statistical filtering to eliminate noise. Ground point segmentation uses the RANSAC algorithm to fit a plane, and non-ground points are clustered as obstacles. Obstacle height information is obtained from the Z-axis histogram of the point cloud. LiDAR calibration uses a calibration board, and intrinsic and extrinsic parameters are periodically checked. Point cloud data timestamps are synchronized with the system clock. Supplementary data acquisition from the visual sensor is also included. The visual sensor uses a binocular camera to calculate depth at a baseline distance of 20cm. Depth calculation is based on a disparity map, and texture features are extracted and fused with color and edge data. The system utilizes YOLOv4 for dynamic target detection, achieving a real-time inference speed of 30fps. Color marking identifies specific RGB values, and extended Kalman filtering smooths the trajectory for dynamic target motion estimation. In the data fusion section, the coordinate transformation matrix is ​​obtained through hand-eye calibration, and nonlinear optimization is used for transformation error compensation. The origin of the local environment coordinate system is updated as the robot moves, and the coordinate transformation chain is maintained consistently. Multidimensional spatial data storage uses database indexes for fast retrieval. Data stream processing pipelines are parallelized to improve throughput. Environmental perception device power management dynamically adjusts sensor frequencies, and an energy-saving mode reduces the sampling rate during idle periods. The entire system is integrated and tested to verify data accuracy, ensuring comprehensive multidimensional spatial data acquisition covering the entire working environment.

[0031] Details of the visual sensor feature extraction algorithm: The Local Binary Pattern (LBP) algorithm calculates the LBP value of each pixel to form a texture map. The map is segmented into blocks, and histograms are used as feature vectors. Ground accessibility classification uses a logistic regression model. The model training data includes various ground types. The dynamic target color marking recognition threshold is configurable to adapt to different lighting conditions. Motion vector calculation uses... Optical flow method, pyramid optical flow improves robustness to large motions. During data fusion, feature point matching is used to align point clouds and images. SIFT feature point extraction and matching are used, and RANSAC is used to remove mismatches. The construction of the local environment coordinate system includes coordinate origin drift correction, which is achieved through loop closure detection. The temporal consistency of multidimensional spatial data is maintained using sliding window optimization. The data output format of the environmental sensing device is standardized. The JSON format includes timestamps, sensor IDs, and data volumes. An anomaly handling mechanism is implemented in the data acquisition module, switching to a backup source when a sensor fails. The entire implementation method has been verified through simulation to ensure adaptability to real-world environments.

[0032] The LiDAR point cloud data processing workflow includes: point cloud downsampling using a voxel grid with a 5cm grid size to preserve shape features; obstacle clustering with an Euclidean distance threshold of 10cm; post-cluster bounding box calculation and alignment; geometric contour simplification using a convex hull algorithm; and height information statistics including minimum and maximum Z values. The visual sensor image processing workflow includes: image denoising using Gaussian filtering; texture feature extraction using multi-channel LBP; color space conversion from RGB to HSV; dynamic object detection; background model update with adaptive learning rate; data fusion; coordinate transformation; and hand-eye calibration. The method employs quaternion representation of the transformation matrix, aligns the local environment coordinate system with the robot chassis coordinate system, uses lossless encoding for multidimensional spatial data compression, achieves real-time data acquisition latency of less than 100ms, implements an automated calibration script for the environmental perception device, and evaluates calibration accuracy with a reprojection error of less than 1 pixel. The entire system is deployed on a robot embedded platform, with CPU load optimized for multi-threaded processing. The acquisition of multidimensional spatial data provides reliable input for path planning. The environmental perception device efficiently and reliably acquires multidimensional spatial data, with LiDAR and vision sensors working collaboratively, achieving accurate data fusion, and maintaining a stable local environment coordinate system, thus supporting the robot's autonomous navigation requirements.

[0033] Example 2: See Figure 3 Multidimensional spatial data comes from local environmental coordinate system data acquired by environmental perception devices. Spatial discretization processing divides the local environmental coordinate system into several grid cells of equal size. The size of the grid cell is set as a multiple of the robot's minimum turning radius. Each grid cell records the obstacle occupancy probability and the ground friction coefficient. The obstacle occupancy probability is calculated by the point density of the LiDAR point cloud data. The higher the point density, the closer the probability value is to 1. The ground friction coefficient is obtained by mapping the ground texture features extracted by the visual sensor to a preset friction coefficient table. Different texture types correspond to different coefficient values. The grid cell attribute update frequency is synchronized with the sensor data refresh rate.

[0034] Adjacent free grid cells are aggregated to form super nodes based on the connectivity between grid cells. The connectivity is determined using the eight-neighbor connectivity rule. Free grid cells are defined as cells with an obstacle occupancy probability of less than 0.3. The super node generation process uses a region growing algorithm, traversing adjacent free grid cells from the seed grid to the boundary. Each super node records the geometric center coordinates of its member grid cells. The super node represents the topological center of the passable area. When establishing connections between super nodes, Euclidean distance is calculated. Euclidean distance is calculated based on the straight-line distance formula of the super node center coordinates. The ground friction coefficient is involved in the calculation of the area passability weight. The area passability weight formula integrates Euclidean distance and average friction coefficient, and the weight value is proportional to the distance and friction coefficient. The topology map has a hierarchical structure consisting of a grid layer and a super node layer. The grid layer stores the original environmental data, while the super node layer abstracts the passable areas. Node connections are established based on the adjacency of super nodes, and adjacency is determined by the boundary distance threshold of super nodes. Area passability weights are assigned to the connecting edges, and the weight matrix is ​​stored in memory for path planning. The spatial discretization process is dynamically updated, and the grid attributes are recalculated when sensors detect changes in the environment. The topology map update mechanism ensures the timeliness of the environmental representation.

[0035] In the topology map, a starting node and a target node are selected. The starting node is selected by mapping the robot's current position provided by the positioning system to the nearest super node. The target node is selected by matching the super node with the nearest neighbor algorithm based on the coordinates input by the user. The path planning model is established based on manifold learning technology. The diffusion mapping algorithm is used to process the high-dimensional spatial relationships of the super nodes. The diffusion mapping algorithm constructs a node similarity matrix, and the matrix elements are based on the connection weights between nodes. The eigenvalue decomposition of the similarity matrix is ​​used to obtain the principal feature vectors. The low-dimensional manifold embedding representation is achieved by retaining the first k feature vectors. After dimensionality reduction, the node coordinates maintain the topological relationship in the manifold space.

[0036] In low-dimensional manifold space, geodesic distance is used instead of Euclidean distance. Geodesic distance calculation is based on finding the shortest path on the manifold. Dijkstra's algorithm is used to search for the shortest path on the manifold graph structure. The connection cost is redefined as a weighted sum of geodesic distance and regional access weights. The connection cost updates the optimal path candidate set. The optimal path candidate set is generated using the k-shortest path algorithm, which iteratively finds the k paths with the minimum cost. The path planning model outputs a candidate set containing multiple feasible paths, satisfying the diversity requirement. Details of high-dimensional spatial relationship processing for supernodes are provided. The node similarity matrix uses a hot kernel function to define similarity, with hot kernel parameters controlling the local neighborhood range. Eigenvalue decomposition employs... The algorithm improves the efficiency of large matrix calculations. The low-dimensional manifold embedding dimension k is adaptively selected according to the number of nodes, usually 3-5 dimensions. The node coordinates in the manifold space are used for visualization analysis. When calculating the geodesic distance, a manifold graph is constructed. The edge weights in the graph are based on the Euclidean distance in the embedding space. The graph structure is preprocessed before running the Dijkstra algorithm to accelerate the query.

[0037] Real-time performance optimization of the path planning model: The diffusion mapping algorithm adopts an incremental update mechanism, which only updates the embedded coordinates of the affected nodes when the topology map changes locally. The geodesic distance calculation caches the distance of commonly used node pairs to reduce redundant calculations. The k-shortest path algorithm uses the Yen algorithm to ensure that the path is not repeated. The candidate set size k is set according to the complexity of the environment. A larger k value in a complex environment increases the diversity of paths. The path planning model is integrated into the robot control system, and the running frequency is matched with the planning cycle.

[0038] The maintenance mechanism of the topology map includes persistent storage of grid cell attributes to support power failure recovery, dynamic merging and splitting of super nodes to handle environmental changes, merging operations when a new connecting channel appears between adjacent super nodes, splitting operations triggered when an obstacle appears within a super node, and learning and updating of area access weights by adjusting weight coefficients based on the robot's actual access experience, with machine learning models recording historical access data to optimize weight formulas.

[0039] The mathematical implementation of the diffusion mapping algorithm includes: similarity matrix normalization to ensure matrix symmetry and positive definiteness; eigenvector standardization to unit vectors; and scaling of embedded coordinates to maintain relative distance proportions. The manifold-learned space is compatible with traditional path planning algorithms. The A* algorithm directly applies heuristic functions based on embedded coordinates in the manifold space, naturally improving path smoothness due to the manifold structure's consideration of environmental topology. Detailed calculation of node connection costs is performed, incorporating robot kinematic parameters into the cost function. Turning radius constraints are converted into direction change penalties, and the impact of dynamic obstacles is mitigated by temporarily adjusting connection costs. A cost update message broadcast mechanism ensures timely planner response. The optimal path candidate set evaluation metrics include path length smoothness and safety, and a multi-objective optimization algorithm balances various metrics. The path planning model interfaces with upper-level task planning, supporting multi-objective point navigation through target node sequence processing. Path replanning triggers are configured with various conditions, such as deviation exceeding limits or drastic environmental changes. Model parameter configuration files facilitate adaptation to different scenarios, and computational resource allocation is dynamically adjusted. In complex environments, the priority of planning threads is increased. Spatial discretization generates a hierarchical topology map, and manifold learning improves path planning quality, enabling the system to adapt to dynamic environmental requirements.

[0040] Example 3: The optimal path candidate set comes from the output of a path planning model based on manifold learning. Key node sequence extraction requires time series analysis of the optimal path candidate set. The time series analysis checks the continuity of the timestamps of the node sequences. Motion continuity constraints include velocity change rate and acceleration limits. The velocity change rate threshold is set to 0.2 meters per second, and the acceleration limit is set to 0.5 meters per square second based on the robot's mass. The node combination screening algorithm traverses the candidate paths and removes node sequences that do not meet the motion constraints. Dynamic target motion information predicts the path conflict risk within the future time window. The prediction time window length is set to 3 seconds. The Kalman filter algorithm estimates the position and velocity of the dynamic target. The path conflict risk calculation is based on the intersection analysis of the robot path and the predicted trajectory of the dynamic target. The collision probability threshold is set to 0.3. Node sequences with a collision probability are removed from the candidate set.

[0041] The remaining node sequences are sorted in ascending order by regional access weight, which is derived from the supernode connection edge weights in the topology map. The sorting algorithm uses quicksort. The first K sequences are selected as input for the initial navigation trajectory. The value of K is dynamically adjusted based on environmental complexity: K=5 for complex environments and K=2 for simple environments. The initial navigation trajectory generates a pathpoint sequence containing position coordinates and expected speed. Dynamic path optimization monitors the dynamic obstacle positions updated by the environmental perception device in real time. These positions are detected through a fusion of visual sensors and LiDAR, with an update frequency of 5Hz. The access weight adjustment for affected grid cells is based on obstacle proximity, calculated as the reciprocal of the distance between the obstacle and the grid center. The weight adjustment coefficient is proportional to the proximity. A rolling temporal optimization strategy recalculates the local paths of the key node sequences at fixed time intervals (0.5 seconds). Local path recalculation is based on the latest topology map connectivity data. The path search uses the A* algorithm, with the heuristic function based on the Euclidean distance from the current node to the target. When optimizing local paths in the rolling temporal domain, the objective function minimizes the path length and the cost derived from the access weights. The passage weight is transformed into an equivalent cost with length dimensions by multiplying it by a normalization factor with length dimensions. The objective function is defined as follows:

[0042] in: This represents the total path cost from the current node to the local target node, measured in units of length. This represents a sequence consisting of N path nodes. This represents the i-th node in the sequence; Indicates adjacent nodes and The Euclidean distance between them is measured in units of length. Indicates the connection node and The region pass weight of the edge is a dimensionless scalar. It is a reference length (taken as 1 meter) used to transform dimensionless weights into cost terms with length dimensions. and A dimensionless coefficient used to balance the weights of path length and travel difficulty.

[0043] When a sudden obstacle is detected, a path replanning mechanism is triggered. This obstacle detection is achieved through anomaly identification in sensor data. The path replanning mechanism then comprehensively replans the path from the current location to the target point, using the same manifold-learned path planning model for alternative trajectory generation. The time series analysis is implemented using a sliding window method with a window size of 10 consecutive nodes. Motion continuity constraint verification calculates the velocity difference and acceleration values ​​between adjacent nodes, with a velocity difference threshold of 0.3 m / s² and an acceleration threshold of 0.5 m / s². Dynamic target motion prediction uses a linear dynamic system model, with model parameters learned from historical observation data. Path conflict risk calculation employs the Monte Carlo method to simulate multiple possible trajectories, and collision probability statistics are used to determine the trajectory intersection ratio. The node sequence sorting algorithm is optimized, using heap sort to improve efficiency for large-scale data, and regional traffic weight normalization avoids dimensional differences.

[0044] The initial navigation trajectory pathpoint interpolation uses cubic spline curves with an interpolation density of 10 points per meter. Trajectory smoothness is evaluated based on the rate of curvature change. Dynamic obstacle position tracking uses a multi-target tracking algorithm, and data association uses the Hungarian algorithm. Weight adjustment is guaranteed to be completed within 100 milliseconds. The path replanning mechanism is triggered when sensors detect new obstacles or the robot deviates from the planned path by more than 0.2 meters. The replanning process prioritizes the use of cached topology maps to reduce computational latency.

[0045] During the key node sequence extraction process, node importance is evaluated based on the centrality measure of a node in the path, and nodes with high centrality are retained as key nodes. Dynamic path optimization introduces safety constraints, ensuring that the minimum distance between the path and obstacles is maintained at more than 0.5 meters. The optimization algorithm uses gradient descent to find the optimal path.

[0046] Sensor data time alignment uses interpolation methods to ensure the accuracy of dynamic obstacle position estimation. The computational complexity of the entire implementation process is analyzed, with time series analysis having a time complexity of O(n). In the dynamic path optimization system, the O(n) time complexity of time series analysis is derived based on the linear traversal characteristic of the node sequence. The object of time series analysis is the node sequence extracted from the optimal path candidate set, assuming the sequence contains n nodes. The analysis process uses a sliding window method, with a fixed window size of 10 nodes. The algorithm processes the node combinations within each window sequentially, and each node is processed only once as the window slides along the sequence.

[0047] Motion continuity constraint verification calculates the velocity difference and acceleration values ​​between adjacent nodes. These calculations are performed within a fixed-size window, and the number of operations is linearly related to the number of nodes, n. Dynamic target motion prediction uses a linear dynamic system model, which performs constant-time operations on each node. Therefore, the overall time complexity is proportional to the number of nodes, n, and is denoted as O(n). Dynamic optimization O(n) 2 The time complexity of the dynamic optimization process is O(n^2). 2 The derivation of this approach stems from its nested loop structure. Dynamic optimization requires handling the interactions between any two nodes in the node sequence; when an environmental change is detected, the algorithm needs to recalculate the connectivity of all node pairs within the affected region. For a sequence containing n nodes, the worst-case scenario requires checking... For a given number of node pairs, the computational complexity increases quadratically with n. While the rolling time-domain optimization strategy limits the number of nodes optimized per iteration, the number of nodes that still needs to be processed in complex environments can still be large, leading to a computational complexity of O(n^2). 2 This system meets real-time requirements. The interface design for dynamic path optimization and motion control models utilizes shared memory for pathpoint sequence transmission, with data format including position, velocity, and timestamp. Anomaly handling mechanisms include degradation strategies for sensor failures, such as continuing operation using only LiDAR data. Optimization algorithm parameters are adaptively adjusted online based on dynamic changes in robot battery level and computational load. Through these implementations, key node sequence extraction and dynamic path optimization enable safe and efficient robot navigation.

[0048] Example 4: The initial navigation trajectory comes from the path point sequence output by the dynamic path optimization module. The motion control model analyzes the curvature change points and velocity constraints of the initial navigation trajectory. The curvature change point identification algorithm is based on the turning angle change rate of the trajectory points. The velocity constraints are derived from the robot dynamic parameters. The piecewise polynomial fitting function adopts the cubic spline interpolation method. Each trajectory segment is represented by a cubic polynomial. The polynomial coefficients are solved through boundary conditions.

[0049] Transition points are inserted into the piecewise polynomial fitting function, located at the junctions of polynomial segments. Robot acceleration continuity is ensured by maintaining the continuity of the second derivative of the function. The position of the transition points is adaptively adjusted based on the rate of curvature change. Feasibility of the transition points is verified based on the steering angle limitations of the wheeled drive mechanism. The steering angle limitation is based on the robot's mechanical structure parameters. The verification algorithm calculates the theoretical steering angle for each connection point and eliminates trajectory segments exceeding the maximum steering angle. The motion control model smooths the trajectory based on robot kinematic constraints, establishing the robot chassis kinematic equations. These equations include wheelbase and wheel diameter parameters. The maximum linear velocity is obtained from motor performance parameters, and the angular velocity limitation is based on the steering mechanism characteristics. Acceleration and deceleration limitations consider robot inertia and motor torque. The execution path is discretized into dense waypoints with a waypoint spacing of 0.1 meters. The differences in motion parameters between adjacent waypoints are calculated, including linear velocity and angular velocity differences. If the difference in motion parameters exceeds a threshold, transition waypoints are generated using Bézier curve interpolation. The threshold is set according to the motion stability requirements. The Bézier curve control points are calculated based on the original waypoint positions. The interpolation process is iterated until all waypoints satisfy the kinematic constraints. The output of the motion control model is a smooth trajectory point sequence containing position, velocity, and acceleration information, which is directly sent to the drive controller for execution.

[0050] The trajectory analysis stage processes the data structure of the initial navigation trajectory. The path point sequence includes Cartesian coordinates and orientation angles. Curvature calculation uses a geometric method based on adjacent three points, and velocity constraints consider the limitation of path curvature on the maximum speed. Cubic splines are chosen as the piecewise polynomial fitting function due to their balance between computational efficiency and smoothness. Solving the polynomial coefficients is transformed into a tridiagonal linear equation system, which is solved efficiently using the pursuit method. The transition point insertion strategy considers motion smoothness, and the number of transition points is dynamically adjusted according to the trajectory complexity.

[0051] Referring to Table 1, the feasibility verification module accesses the robot mechanism parameter database. The steering angle limit is related to the current speed, with stricter limits at high speeds. Kinematic equations are established based on a differential drive model, including the conversion relationship between linear velocity and left / right wheel speeds. Acceleration and deceleration limits are stored in a lookup table, with different limits corresponding to different speeds. Waypoint discretization maintains path shape accuracy, and the central difference method is used to improve accuracy in calculating motion parameter differences. Bézier curve interpolation uses a quadratic curve form, with three control points defining a smooth transition curve. Motion parameters are recalculated after interpolation. The entire smoothing process is timed within 50 milliseconds, meeting control cycle requirements. The motion control model communicates with the underlying driver via a CAN bus, with data including target position and velocity commands. Tracking errors are monitored in real-time during trajectory execution; replanning is triggered when the error exceeds a threshold. Motion control model parameters can be adjusted online to adapt to different load conditions. Through the above implementation, the initial navigation trajectory is transformed into an executable smooth path, ensuring smooth and accurate robot movement.

[0052] Table 1: Transition Connection Point Parameter Configuration

[0053] The specific implementation of the piecewise polynomial fitting function involves a trajectory segmentation strategy. Segmentation points are selected based on curvature extrema, and the polynomial parameters within each segment are solved independently. A transition point insertion algorithm detects acceleration discontinuities, inserts new points at discontinuities, and eliminates abrupt changes by adjusting the polynomial coefficients. The steering angle limit for the wheel drive mechanism is determined by querying the mechanism parameter table, and the limit value is dynamically adjusted for different speeds. Kinematic equation modeling considers wheel-ground friction, and the maximum linear velocity is appropriately reduced on slippery surfaces. The waypoint discretization process maintains a constant path length, and dense waypoints improve control accuracy. The motion parameter difference calculation includes velocity and acceleration vectors, and the threshold setting is based on ride comfort requirements. The Bézier curve control point calculation uses the de Castelli algorithm to generate smooth transition curves.

[0054] After implementing the motion control model, the robot trajectory tracking error test results met the accuracy requirements, and the acceleration continuity was better than that of the unsmoothed trajectory. The computational complexity of the entire processing flow is linear with the number of waypoints, making it suitable for real-time operation on embedded platforms. Model parameters can be adjusted via configuration files to adapt to different robot models.

[0055] Example 5: Using an autonomous robot inspecting a large underground parking garage as an example, the robot's task is to start from the security control room in the southeast corner of level B2, inspect all 120 parking spaces along a predetermined route, and return to the starting point. The total path length is 1.8 kilometers. The path planning module sets 22 checkpoints along this path, with the spacing dynamically adjusted according to environmental characteristics. In areas with open views, the spacing is set to 100 meters, while in areas with dense pillars and complex structures, the spacing is shortened to 50 meters. The checkpoint triggering mechanism is tightly coupled with the robot's positioning system. When the robot confirms that it has reached within 3 meters of the checkpoint coordinates through wheeled odometer and laser SLAM fusion positioning, it immediately triggers the local data update process of the environmental perception device. The laser radar of the environmental perception device scans the surrounding environment at a frequency of 10Hz, and the visual sensor simultaneously acquires 5 frames of high-definition images. The acquired raw data is compared with the pre-stored baseline data of the area in the topology map. The comparison algorithm extracts the feature descriptors of the current point cloud and matches them with the features stored in the map to calculate the difference between the actual environmental data and the topology map.

[0056] In a specific scenario, the robot approaches checkpoint CT-07, located near a bend at the intersection of parking areas B and C. Data from the environmental perception device shows a temporary fire extinguisher cabinet has been added between two pillars previously marked on the map. The difference calculation module iteratively registers the point cloud data to the nearest point, but the registration error reaches 0.25 meters, exceeding the 0.15-meter tolerance threshold. The system immediately issues a pause command, and the robot drive system executes a smooth braking procedure, stopping smoothly from a speed of 1.2 meters per second within 0.5 seconds, simultaneously triggering a map correction process. After the map correction process begins, the robot re-acquires LiDAR point cloud data for its current location. The LiDAR performs a 360-degree panoramic scan, acquiring a 3D point cloud with a density of 2000 points per square meter. The visual sensors simultaneously acquire ground texture features, focusing on capturing visual features around the newly added obstacle. This data is used by a sensor fusion algorithm to generate a corrected local environmental coordinate system. The new coordinate system uses the robot's current stopping position as the origin, accurately calibrating the boundary coordinates of the newly added obstacle.

[0057] Based on the corrected environmental data, the topology map update module adjusts the supernode connections in the affected areas. The supernode link originally connecting areas B and C needs to have its connectivity recalculated due to the added obstacle. After detecting the blocked original path, the path planning algorithm searches for alternative paths in adjacent areas and generates new supernode connections. The area access weight calculation module then updates the weight values ​​of the affected areas; the grid weight around the added obstacle is adjusted from 1.0 to 2.5, and the weights of adjacent channels are dynamically adjusted accordingly. Based on the updated topology map, the path replanning module regenerates the latter half of the execution path starting from the current checkpoint. The replanning algorithm uses an incremental planning strategy, retaining the validity of already traveled paths and only replanning untraveled path segments. The newly planned path bypasses the added obstacle and selects adjacent channels to continue the inspection task. The entire map correction process, from triggering to completion, takes 3.2 seconds, with the robot's pause time kept within acceptable limits.

[0058] The checkpoint placement strategy takes into account the characteristics of the parking lot environment, selecting each checkpoint location in an area with good signal coverage to ensure reliable data transmission. The difference calculation employs a multi-feature fusion method, including point cloud matching degree, visual feature similarity, and obstacle position deviation. The weights of the comprehensive difference calculation formula have been validated through extensive experiments. The tolerance threshold is set according to the parking lot inspection accuracy requirements; a tolerance value of 0.15 meters ensures timely detection of environmental changes while avoiding false triggers due to sensor noise. The point cloud registration algorithm in the map correction process uses a normal distribution transformation method, achieving centimeter-level registration accuracy. Visual feature extraction uses an improved ORB feature descriptor, exhibiting strong robustness to changes in parking lot lighting. The topology map update adopts an incremental update mechanism, modifying only the map data of changed areas to improve update efficiency. The passage weight adjustment considers obstacle type; the weight adjustment range for temporary obstacles is smaller than that for fixed obstacles, facilitating rapid adjustment during subsequent environmental recovery.

[0059] The path replanning module employs a rolling temporal planning strategy, dynamically adjusting the planning field of view based on environmental complexity. During new path generation, the motion control model simultaneously performs trajectory smoothing to ensure the replanned path meets the robot's kinematic constraints. The entire system utilizes a multi-threaded parallel processing architecture, with perception, planning, and control modules working collaboratively to minimize robot downtime. During the parking lot inspection task, the robot triggered four map correction processes, addressing environmental changes such as temporarily parked vehicles, newly added fire extinguishers, and construction barriers. After each map correction, the system automatically records environmental change information to optimize subsequent checkpoint setting strategies and difference calculation parameters. Through this adaptive mechanism, the robot can continuously adapt to dynamically changing parking lot environments without relying on prior map updates, enabling it to complete long-term autonomous inspection tasks.

[0060] The effectiveness of the checkpoint mechanism was verified through continuous operation. The robot was able to detect changes in the parking lot layout in a timely manner and accurately trigger the map correction process. The corrected path maintained its optimality, ensuring the efficient completion of the inspection task. The reliability of the map correction process ensured the long-term stable operation of the robot in complex and dynamic environments, providing a solid technical foundation for the autonomous inspection system. The entire implementation process demonstrated the practical value of the checkpoint mechanism and map correction process in real-world scenarios, reflecting the system's adaptability and robustness to dynamic environments.

[0061] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for route searching of an automatic route-searching intelligent robot, characterized in that, The method comprises the following steps: acquiring multi-dimensional space data of the environment where the robot is currently located by an environmental perception device, the multi-dimensional space data comprising obstacle distribution information, ground feature information and dynamic target motion information; performing spatial discretization processing on the multi-dimensional space data to generate a topological map with a hierarchical structure, the topological map comprising node connection relationships and region passing weights; establishing a path planning model based on manifold learning according to the node connection relationships of the topological map, the path planning model determining an optimal path candidate set by calculating the manifold distance between adjacent nodes; extracting a key node sequence in the optimal path candidate set, and performing dynamic path optimization in combination with the region passing weights to generate a preliminary navigation trajectory; inputting the preliminary navigation trajectory into a motion control model, the motion control model performing smoothing processing on the trajectory according to the kinematic constraints of the robot, and outputting a final execution path. 2.The route-searching method of the automatic route-searching intelligent robot according to claim 1, wherein, The step of acquiring multi-dimensional space data of the environment where the robot is currently located by an environmental perception device comprises the following steps: collecting three-dimensional point cloud data of the environment by using a laser radar, the three-dimensional point cloud data comprising the geometric profile and height information of obstacles; extracting ground texture features and color markers of dynamic targets by using a visual sensor, the ground texture features being used to distinguish passable regions from impassable regions; fusing the three-dimensional point cloud data and the ground texture features to construct a local environment coordinate system with the current position of the robot as the origin, the local environment coordinate system being used to unify the data space reference. 3.The route-searching method of the automatic route-searching intelligent robot according to claim 2, wherein, The step of performing spatial discretization processing on the multi-dimensional space data to generate a topological map with a hierarchical structure comprises the following steps: dividing the local environment coordinate system into a plurality of equal-size grid cells, each grid cell recording obstacle occupancy probability and ground friction coefficient; aggregating adjacent idle grid cells to form super nodes according to the connectivity between the grid cells, the super nodes representing the topological centers of passable regions; calculating region passing weights based on the Euclidean distances between the super nodes and the ground friction coefficients, the region passing weights being used to quantify the path selection cost. 4.The route-searching method of the automatic route-searching intelligent robot according to claim 3, wherein, The step of establishing a path planning model based on manifold learning comprises the following steps: selecting a start node and a target node in the topological map, the start node corresponding to the current position of the robot, and the target node being input by a user or specified by a task planning system; performing dimension reduction processing on the high-dimensional space relationship of the super nodes by using a diffusion mapping algorithm to generate a low-dimensional manifold embedding representation; in the low-dimensional manifold space, redefining the connection cost between nodes by using geodesic distance instead of Euclidean distance, the connection cost being used to update the optimal path candidate set. 5.The route-searching method of the automatic route-searching intelligent robot according to claim 4, wherein, The step of extracting a key node sequence in the optimal path candidate set comprises the following steps: performing time series analysis on the optimal path candidate set to filter node combinations that satisfy the motion continuity constraint; predicting the path conflict risk in a future time window according to the dynamic target motion information, and eliminating node sequences that have a collision probability; arranging the remaining node sequences in ascending order of region passing weights, and selecting the first K sequences as the input of the preliminary navigation trajectory. 6.The route-searching method of the automatic route-searching intelligent robot according to claim 5, wherein, The step of performing dynamic path optimization in combination with the region passing weights comprises the following steps: Real-time monitoring of dynamic obstacle positions updated by the environmental perception device, adjusting the passing weight of the affected grid cells; Adopting a receding horizon optimization strategy to recalculate the local path of the key node sequence at fixed time intervals; When a sudden obstacle is detected, triggering the path re-planning mechanism, which generates an alternative trajectory based on the latest topological map. 7.The route-searching method of the automatic route-searching intelligent robot according to claim 6, wherein, The preliminary navigation trajectory is input into a motion control model, including: Analyzing the curvature change points and speed constraint conditions of the preliminary navigation trajectory to generate a segmented polynomial fitting function; Inserting transition joints in the segmented polynomial fitting function to ensure continuous acceleration of the robot; According to the steering angle limit of the wheeled driving mechanism, the feasibility of the transition joint is verified, and the unexecutable trajectory segment is removed. 8.The route-searching method of the automatic route-searching intelligent robot according to claim 7, wherein, The motion control model smoothes the trajectory according to the kinematic constraints of the robot, including: Establishing the kinematic equation of the robot chassis, which contains the maximum linear speed, angular velocity and acceleration / deceleration limit; Discretizing the final execution path into dense waypoints and calculating the motion parameter difference between adjacent waypoints; If the motion parameter difference exceeds the threshold, a transition waypoint is generated using the Bezier curve interpolation method until all waypoints meet the kinematic constraints. 9.The route-searching method of the automatic route-searching intelligent robot according to claim 8, wherein, Also including: Setting multiple checkpoints on the final execution path, which are used to trigger local data updates of the environmental perception device; When the robot reaches the checkpoint, comparing the difference between the actual environment data and the topological map, if the difference exceeds the tolerance, suspending the execution and starting the map correction process. 10.The route-searching method of the automatic route-searching intelligent robot according to claim 9, wherein, The map correction process includes: Recollecting the laser radar point cloud data and visual texture features of the current position to generate a revised local environment coordinate system; Updating the super node connection relationship of the topological map according to the correction result, and recalculating the passing weight of the affected area; Based on the updated topological map, the second half of the final execution path is regenerated from the current checkpoint.

Citation Information

Cited By

  • Mobile robot dynamic path planning method suitable for farmland environment

    CN121677742A

  • Cleaning robot obstacle recognition method based on visual saliency detection

    CN121747080A

  • Control method and system based on medical transportation robot

    CN121766874A

  • AMR robot optimal path planning device and method based on Internet of Things

    CN121898439A