An adaptive mapless navigation method and system for non-flat terrain

Abstract

The application provides a self-adaptive map-free navigation method and system for uneven terrain, and relates to the field of mobile robots, and the following steps are iteratively performed until the final target point is reached based on the current position of the robot: positioning the current position based on multi-source sensor data of the robot, and constructing a local map; selecting an optimal boundary vertex sequence from the local map according to the current positioning result and the local map, and then obtaining an optimal sub-target point; performing passability analysis and collision detection based on the optimal sub-target point, and generating a path to the sub-target point; and predicting a trajectory in real time according to the current generated path and the local map, and driving the robot to move to the sub-target point along the predicted trajectory; and the application enables the mobile robot to adaptively navigate on uneven terrain under the condition of no prior map, and improves the adaptability and convenience of the autonomous navigation system.

Description

Technical Field

[0001] This invention belongs to the field of mobile robots, and in particular relates to an adaptive mapless navigation method and system for non-flat terrain. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Robotics research is a topic of continuous interest to scholars both domestically and internationally, particularly the issue of mapless navigation for robots in uneven terrain, which has long been a technological bottleneck in the robotics industry. Mobile robots, with their advantages of flexible movement, ease of operation, and robustness, are widely used in fields such as medicine, military, aerospace, and logistics. Traditional navigation systems typically rely on pre-built map data for guidance. However, in large-scale environments such as mountains, forests, or wilderness, map data may be incomplete and map construction may be difficult. In such cases, adaptive mapless navigation systems enable robots to utilize peripheral sensors and local perception information to guide them through a series of sub-target points in complex, uneven terrain, gradually reaching the final target point and achieving navigation functionality, thus greatly improving the convenience of robot navigation.

[0004] In existing technical solutions, map-free navigation technology for uneven terrain mainly faces the following technical challenges:

[0005] 1. Complex terrain accessibility analysis: Non-flat terrain contains various terrain features and obstacles, such as mountains, steep slopes, and trees. How to extract navigation-related passable areas from sensor data and limited local maps, and perform real-time analysis and modeling, is a key problem that urgently needs to be solved.

[0006] 2. Obstacle identification and safety assessment: On uneven terrain, obstacle identification differs from that on flat terrain. Changes in the shape and position of obstacles often affect the safety of the robot when passing through, and there may be occlusions or blind spots. Therefore, accurately detecting and identifying dangerous obstacles and making real-time decision-making and planning is a challenging task.

[0007] 3. Real-time path planning and decision-making: How to make efficient decisions and plan paths in real time without prior maps, while taking into account changes in terrain, the presence of obstacles, and the requirements of the navigation target, is a complex problem.

[0008] Therefore, existing mapless navigation technologies for uneven terrain suffer from difficulties in accurately identifying obstacles and accurately modeling terrain, resulting in insufficient navigation efficiency and accuracy. Summary of the Invention

[0009] To overcome the shortcomings of the prior art, the present invention provides an adaptive mapless navigation method and system for uneven terrain, which enables mobile robots to navigate adaptively on uneven terrain without prior maps, thereby improving the adaptability and convenience of autonomous navigation systems.

[0010] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0011] The first aspect of the present invention provides an adaptive mapless navigation method for non-flat terrain.

[0012] An adaptive map-free navigation method for non-flat terrain iteratively executes the following steps based on the robot's current position until the final target point is reached:

[0013] Based on multi-source sensor data from the robot, the current location is determined and a local map is constructed;

[0014] Based on the current positioning results and local map, adaptive decision-making and real-time planning are performed to generate the next sub-target point and the corresponding path;

[0015] Based on the currently generated path and local map, the trajectory is predicted in real time, driving the robot to move along the predicted trajectory to the sub-target point;

[0016] The adaptive decision-making and real-time planning involve using an adaptive decision network to select the optimal sequence of boundary vertices from the local map, thereby obtaining the optimal sub-target point. The elevation-aware planning tree then performs accessibility analysis and collision detection based on the optimal sub-target point to generate a path to the sub-target point.

[0017] Furthermore, the process of locating the current position specifically involves:

[0018] The robot's pose is calculated based on LiDAR point cloud data and inertial sensor data respectively, and pose fusion is performed to obtain the pose estimate;

[0019] Based on the keyframe sequence in the environment, the pose estimation is optimized to obtain the optimized pose.

[0020] Furthermore, the construction of the local map specifically involves:

[0021] The lidar point cloud data is mapped onto a 3D voxel map, and the 3D voxel map is windowed and projected onto a plane to construct a 2D elevation grid map.

[0022] A bicubic interpolation-based image smoothing algorithm is used to interpolate a two-dimensional elevation map to obtain the final local map.

[0023] Furthermore, the adaptive decision network takes all ordinary vertices and obstacle vertices in the local map and the elevation-aware planning tree as input, performs dimensionality reduction mapping on all input data to form a feature map with two-dimensional image features, extracts features from the feature map, evaluates vertices at the map boundary based on the extracted features, selects the optimal boundary vertex sequence, obtains the sub-target point sequence in space through spatial mapping, arranges the sub-target points according to the optimality metric of the sub-target points, and selects the first sub-target point as the optimal sub-target point for the next step.

[0024] Furthermore, the elevation-aware planning tree includes two structures: a sliding window search tree and a global undirected graph. The sliding window search tree uses the RRT* algorithm to grow in a local map, and its root vertex always moves with the robot's position. During the tree expansion process, the passability of edges is determined by judging the flatness of the ground, and dangerous areas in the map are marked by obstacle vertices.

[0025] Furthermore, the global undirected graph is constructed using a probabilistic route graph algorithm without sampling, which is responsible for recording the sub-target points explored by the robot and the paths leading to these sub-target points. Each vertex of the global undirected graph is provided by an adaptive decision network and a sliding window search tree.

[0026] Furthermore, the predicted trajectory includes trajectory generation and trajectory tracking;

[0027] The trajectory generation aims to minimize time while taking into account navigation efficiency and terrain flatness. It solves a nonlinear optimization problem to obtain the optimal trajectory.

[0028] The trajectory tracking method, based on a differential single-vehicle model and PID control, tracks the predicted trajectory and outputs a control sequence to drive the robot's movement.

[0029] A second aspect of the present invention provides an adaptive mapless navigation system for non-flat terrain.

[0030] An adaptive mapless navigation system for non-flat terrain, employing the method provided in the first aspect for navigation, includes a terrain construction and localization module, an adaptive decision-making and real-time planning module, and a motion control module.

[0031] The terrain construction and localization module is configured to: locate the current position based on the robot's multi-source sensor data and construct a local map;

[0032] The adaptive decision-making and real-time planning module is configured to: make adaptive decisions and real-time plans based on the current positioning results and local map, and generate the next sub-target point and the corresponding path;

[0033] The motion control module is configured to predict the trajectory in real time based on the currently generated path and local map, and drive the robot to move along the predicted trajectory to the sub-target point.

[0034] The adaptive decision-making and real-time planning involve using an adaptive decision network to select the optimal sequence of boundary vertices from the local map, thereby obtaining the optimal sub-target point. The elevation-aware planning tree then performs accessibility analysis and collision detection based on the optimal sub-target point to generate a path to the sub-target point.

[0035] A third aspect of the invention provides a computer-readable storage medium having a program stored thereon that, when executed by a processor, implements the steps of an adaptive mapless navigation method for non-flat terrain as described in the first aspect of the invention.

[0036] A fourth aspect of the present invention provides an electronic device including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of an adaptive mapless navigation method for non-flat terrain as described in the first aspect of the present invention.

[0037] The above one or more technical solutions have the following beneficial effects:

[0038] Based on accumulated multi-frame point cloud data and open-loop positioning results, this invention can perform sliding window-type surface fitting on non-flat terrain environments to obtain local environment maps, effectively solving the problems of map storage and terrain construction in large-scale environments.

[0039] This invention uses a robot as the center to construct an elevation-aware planning tree, which can simultaneously solve path planning and accessibility analysis problems.

[0040] This invention employs a deep learning-based adaptive decision network to select the next sequence of sub-target points, effectively solving the problem of efficient autonomous decision-making under conditions without prior maps.

[0041] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0042] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0043] Figure 1 This is a flowchart of the method in the first embodiment.

[0044] Figure 2This is a system framework diagram for adaptive map-free navigation in non-flat terrain, as shown in the first embodiment.

[0045] Figure 3 This is a framework diagram of the terrain construction and positioning module in the first embodiment;

[0046] Figure 4 The diagram shows the structure of the adaptive decision network based on a deep neural network in the first embodiment.

[0047] Figure 5 This is a schematic diagram of the growth of the elevation-sensing planning tree in the first embodiment;

[0048] Figure 6 This is a schematic diagram of trajectory tracking based on the PID algorithm in the first embodiment;

[0049] The components are: 1. Obstacles; 2. Global undirected graph; 3. Sliding window search tree; 4. Robot; 5. Final target point; 6. Local grid map; 7. Predicted trajectory. Detailed Implementation

[0050] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0051] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0052] Example 1

[0053] In one or more embodiments, an adaptive map-free navigation method for non-flat terrain is disclosed, such as... Figure 1 As shown, based on the robot's current position, the following steps are executed iteratively until the final target point is reached:

[0054] Step S1: Based on the robot's multi-source sensor data, locate the current position and construct a local map;

[0055] Step S2: Based on the current positioning results and local map, perform adaptive decision-making and real-time planning to generate the next sub-target point and corresponding path;

[0056] Step S3: Based on the currently generated path and local map, predict the trajectory in real time and drive the robot to move along the predicted trajectory to the sub-target point;

[0057] The adaptive decision-making and real-time planning involve using an adaptive decision network to select the optimal sequence of boundary vertices from the local map, thereby obtaining the optimal sub-target point. The elevation-aware planning tree then performs accessibility analysis and collision detection based on the optimal sub-target point to generate a path to the sub-target point.

[0058] The following is a detailed description of the implementation process of an adaptive map-free navigation method for non-flat terrain in this embodiment.

[0059] As described in the background section, existing technologies have technical problems. To address these problems, this embodiment discloses a mapless navigation method and system for non-flat terrain in the field. This method can construct and locate a local map of the non-flat terrain based on measurement data from the robot's peripheral sensors. It can simultaneously perform ground accessibility analysis and path planning by constructing an elevation-aware planning tree. An adaptive decision network makes efficient decisions based on the elevation-aware tree and the map, thereby obtaining the next sub-target point sequence and corresponding path. Furthermore, a motion control module drives the robot to complete autonomous navigation. Figure 2 As shown, the three steps are divided into three modules: terrain construction and localization module, adaptive decision-making and real-time planning module, and motion control module. Each module is explained in detail.

[0060] Terrain construction and localization modules, such as Figure 3 As shown, the laser inertial navigation SLAM algorithm and elevation mapping algorithm are used to provide accurate positioning and slide to build local maps.

[0061] This embodiment uses a laser inertial navigation SLAM algorithm to achieve accurate pose estimation, i.e., localization, for the robot. It mainly includes two parts: front-end pose estimation and back-end loop closure optimization, specifically:

[0062] The front-end pose estimation is responsible for fusing the poses calculated from LiDAR point cloud data and inertial sensor data to obtain front-end odometry. At the same time, it extracts key frame sequences from the environment at a certain frequency to provide a basis for back-end loop closure optimization.

[0063] Specifically, the lidar sensor calculates pose using a uniform interpolation method, while the inertial sensor calculates pose using a pre-integration method. The pose calculation results from the two types of sensors are then fused by solving the following nonlinear optimization problem:

[0064]

[0065] Where X represents the robot's posterior pose, β l β i These are the weighting coefficients, ρ[] represents the loss function of the residuals, and r l r i These represent the residuals between the posterior pose and the prior pose provided by the LiDAR and inertial sensors, respectively. By summing the differences between the posterior poses, the robot's current odometer estimate can be obtained.

[0066] Back-end loop closure optimization is responsible for registering robot observation data with historical observation data, thereby reducing global pose estimation error.

[0067] A sliding local map is constructed using an elevation mapping algorithm, specifically as follows:

[0068] Based on the optimized pose, the laser point cloud of the current frame is mapped into a 3D voxel map, and the voxel map is windowed and projected onto a plane to construct a 2D elevation grid map. The value stored in each grid cell of the 2D elevation grid map represents the maximum ground height of the current grid cell position. Finally, the elevation of the 2D elevation map is interpolated by using an image smoothing algorithm based on bicubic interpolation, thereby obtaining the final local grid map and pose estimation.

[0069] The adaptive decision-making and real-time planning module mainly comprises an elevation-aware planning tree and an adaptive decision network. It is responsible for extracting the next sequence of sub-target points based on the current local map and positioning results, and generating paths to these target points. In this invention, the elevation-aware planning tree not only generates paths to target points but also performs terrain accessibility analysis and collision detection during navigation, providing guidance for decision-making. The adaptive decision network is responsible for extracting the next sub-target points based on the elevation-aware planning tree and the current local map, thereby guiding the robot to the sub-target points.

[0070] Adaptive decision networks, such as Figure 4 As shown, it mainly consists of a multilayer perceptron (MLP), convolutional layers (CNN), and a multi-head self-attention module (MHA). Taking all ordinary vertices and obstacle vertices from the local map and the elevation-aware planning tree as input, it first performs a dimensionality reduction mapping operation on all 3D input data to form feature maps with 2D image features. Then, it extracts features from the feature maps through convolutional operations. Figure 4The "⊕" symbol represents the channel stacking operation of the feature map. In solving decision-making problems without map navigation, map boundaries can effectively guide the robot to explore the vicinity of the final target point. The MHA module's role is to enable the adaptive decision network to focus on evaluating vertices at the map boundaries, thereby selecting the optimal sequence of boundary vertices as the decision result. Then, through spatial mapping, decoding is performed to obtain a sequence of sub-target points in space. This sequence is arranged according to the optimality metric of the sub-target points, selecting the first sub-target point as the optimal sub-target point for the next step, and the remaining sub-target points as candidate sub-target points.

[0071] Elevation-aware planning tree, such as Figure 5 As shown, it mainly includes two structures: sliding window search tree and global undirected graph.

[0072] The sliding window search tree grows only on a local map. In this embodiment, the RRT* algorithm is used as the growth method for the sliding window search tree, and its root vertex always moves with the robot's position. The sliding window search tree has real-time terrain perception capabilities. During the tree expansion process, the passability of edges needs to be determined by judging the flatness of the ground. The flatness of the ground can be expressed by formula (1).

[0073]

[0074] Among them, h i Let l represent the ground height at the i-th vertex. i This indicates the length of the current edge.

[0075] When the flatness of the terrain where the edge is located is too large, this edge will not be added to the sliding window search tree. Eventually, the sliding window search tree will grow branches that meet the terrain flatness requirements within the local map area, achieving a preliminary terrain perception effect.

[0076] This embodiment also introduces an obstacle vertex to mark dangerous areas on the map. The location of the obstacle vertex is a location that the robot cannot reach (e.g., steep slopes, cliffs, etc.); the criteria for determining the obstacle vertex are as follows:

[0077]

[0078] in, This represents the total number of times the sampled points attempt to connect to the i-th vertex in the tree during the tree growth process using a sliding window search. This represents the number of times a sample point can successfully connect to the i-th vertex.

[0079] When formula (2) is true, it indicates that the current position of the i-th vertex is not easy to connect with other vertices and is very likely to be in a dangerous area such as a steep slope. Therefore, the i-th vertex is permanently marked as an obstacle vertex until the navigation task ends.

[0080] All obstacle vertices will undergo an expansion operation. During the growth of the sliding window search tree, if a newly added vertex is inside the expanded sphere of an obstacle vertex, the vertex is considered invalid and will not be added to the tree.

[0081] The global undirected graph records the sub-target points explored by the robot and the paths leading to those sub-target points. This allows the robot to re-navigate to other previously explored sub-target points if it gets stuck in a dead end. This embodiment uses a probabilistic route graph algorithm (PRM) without sampling to construct the global undirected graph. Each vertex of the global undirected graph is provided by an adaptive decision network and a sliding window search tree. The vertices in the global undirected graph mainly fall into the following categories:

[0082] (1) Sub-target points output by the adaptive decision network. These vertices help the robot record the sub-target points that have been explored in the past, thus avoiding the robot getting stuck in a dead end.

[0083] (2) Intermediate branch vertices leading to local sub-target points: Among the sub-target points output by the decision network, local sub-target points located in the sliding window search tree are separated, and intermediate branch vertices from the root vertex to these local sub-target points can be searched in the sliding window search tree to provide the robot with a path to the historical sub-target points.

[0084] Based on the sub-target points output by the adaptive decision network, the corresponding paths are queried on the elevation-aware planning tree. The paths are then smoothed using a cubic spline interpolation algorithm, resulting in a smooth path leading to the sub-target points.

[0085] The motion control module is responsible for generating a real-time predicted trajectory based on the currently generated path and local map, and driving the robot to move along the predicted trajectory through a trajectory tracking algorithm. Therefore, it includes a trajectory generation module and a trajectory tracking module.

[0086] The trajectory generation module employs an optimization-based trajectory generation method, the optimization problem of which can be described as:

[0087]

[0088] Where ω1 and ω2 are weight parameters, ΔT is the time interval between two adjacent points on the predicted trajectory, ΔW is the measure of attitude angle change between two adjacent points on the predicted trajectory, r, p, and y are the robot's attitude angles, v, ω, and a are the robot's linear velocity, angular velocity, and acceleration, respectively, and p r p t These represent the robot's current pose and the farthest predicted pose, respectively, using {s1, s2, ..., s...} n} represents the robot's state from the current moment to a future moment, and the optimization objective variable used is s = {s1, s2, ..., s}. n,ΔT1,ΔT2,…,ΔT n By solving the above nonlinear optimization problem, the optimal trajectory that balances navigation efficiency and terrain flatness can be obtained.

[0089] The trajectory tracking module employs a trajectory tracking method based on a differential single-vehicle model and PID control, such as... Figure 6 As shown, the robot's kinematic state model is as follows:

[0090]

[0091] Where x and y are the planar coordinates of the robot relative to the global coordinate origin, and v, θ, and a are the robot's linear velocity, angular velocity, and linear acceleration, respectively.

[0092] Its error variable can be described as the lateral position error of the robot from the point closest to the target trajectory, that is:

[0093] e = dsin(β) e (6)

[0094] Where e represents the lateral error of the robot's current position relative to the expected position of the predicted trajectory. d, β e These represent the Euclidean distance and offset angle of the robot's current position relative to the desired position on the predicted trajectory, respectively.

[0095] By employing PID position control to eliminate lateral position errors, the robot can accurately track the predicted trajectory, output a control sequence to drive the robot's movement, and thus complete the robot's motion control under given path conditions.

[0096] Example 2

[0097] In one or more embodiments, an adaptive mapless navigation system for non-flat terrain is disclosed, which uses the method disclosed in Embodiment 1 for navigation, including a terrain construction and localization module, an adaptive decision-making and real-time planning module, and a motion control module:

[0098] The terrain construction and localization module is configured to: locate the current position based on the robot's multi-source sensor data and construct a local map;

[0099] The adaptive decision-making and real-time planning module is configured to: make adaptive decisions and real-time plans based on the current positioning results and local map, and generate the next sub-target point and the corresponding path;

[0100] The motion control module is configured to predict the trajectory in real time based on the currently generated path and local map, and drive the robot to move along the predicted trajectory to the sub-target point.

[0101] The adaptive decision-making and real-time planning involve using an adaptive decision network to select the optimal sequence of boundary vertices from the local map, thereby obtaining the optimal sub-target point. The elevation-aware planning tree then performs accessibility analysis and collision detection based on the optimal sub-target point to generate a path to the sub-target point.

[0102] Example 3

[0103] The purpose of this embodiment is to provide a computer-readable storage medium.

[0104] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of an adaptive mapless navigation method for non-flat terrain as described in Embodiment 1 of this disclosure.

[0105] Example 4

[0106] The purpose of this embodiment is to provide an electronic device.

[0107] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of an adaptive mapless navigation method for non-flat terrain as described in Embodiment 1 of this disclosure.

[0108] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An adaptive mapless navigation method for non-flat terrain, characterized in that, Based on the robot's current position, iteratively execute the following steps until the final target point is reached: Based on multi-source sensor data from the robot, the current location is determined and a local map is constructed; Based on the current positioning results and local map, adaptive decision-making and real-time planning are performed to generate the next sub-target point and the corresponding path; Based on the currently generated path and local map, the trajectory is predicted in real time, driving the robot to move along the predicted trajectory to the sub-target point; The adaptive decision-making and real-time planning involve using an adaptive decision network to select the optimal sequence of boundary vertices from a local map, thereby obtaining the optimal sub-target point. The elevation-aware planning tree then performs accessibility analysis and collision detection based on the optimal sub-target point to generate a path to the sub-target point. The adaptive decision network takes all ordinary vertices and obstacle vertices in the local map and elevation-aware planning tree as input, performs dimensionality reduction mapping on all input data to form a feature map with two-dimensional image features, extracts features from the feature map, evaluates vertices at the map boundary based on the extracted features, selects the optimal boundary vertex sequence, obtains the sub-target point sequence in space through spatial mapping, arranges the sub-target points according to the optimality metric of the sub-target points, and selects the first sub-target point as the optimal sub-target point for the next step. The elevation-aware planning tree includes two structures: a sliding window search tree and a global undirected graph. The sliding window search tree uses the RRT* algorithm to grow in a local map, and its root vertex always moves with the robot's position. During the tree expansion process, the passability of the edges is determined by judging the flatness of the ground, and dangerous areas in the map are marked by obstacle vertices. The global undirected graph is constructed using a probabilistic route graph algorithm without sampling. It is responsible for recording the sub-target points explored by the robot and the paths leading to these sub-target points. Each vertex of the global undirected graph is provided by an adaptive decision network and a sliding window search tree.

2. The adaptive mapless navigation method for non-flat terrain as described in claim 1, characterized in that, The process of locating the current position specifically involves: The robot's pose is calculated based on LiDAR point cloud data and inertial sensor data respectively, and pose fusion is performed to obtain the pose estimate; Based on the keyframe sequence in the environment, the pose estimation is optimized to obtain the optimized pose.

3. The adaptive mapless navigation method for non-flat terrain as described in claim 1, characterized in that, The construction of the local map specifically involves: The lidar point cloud data is mapped onto a 3D voxel map, and the 3D voxel map is windowed and projected onto a plane to construct a 2D elevation grid map. A bicubic interpolation-based image smoothing algorithm is used to interpolate a two-dimensional elevation map to obtain the final local map.

4. The adaptive mapless navigation method for non-flat terrain as described in claim 1, characterized in that, The predicted trajectory includes trajectory generation and trajectory tracking; The trajectory generation aims to minimize time while taking into account navigation efficiency and terrain flatness. It solves a nonlinear optimization problem to obtain the optimal trajectory. The trajectory tracking method, based on a differential single-vehicle model and PID control, tracks the predicted trajectory and outputs a control sequence to drive the robot's movement.

5. An adaptive mapless navigation system for non-flat terrain, characterized in that, Navigation using the method described in any one of claims 1-4 includes a terrain construction and localization module, an adaptive decision-making and real-time planning module, and a motion control module. The terrain construction and localization module is configured to: locate the current position based on the robot's multi-source sensor data and construct a local map; The adaptive decision-making and real-time planning module is configured to: make adaptive decisions and real-time plans based on the current positioning results and local map, and generate the next sub-target point and the corresponding path; The motion control module is configured to predict the trajectory in real time based on the currently generated path and local map, and drive the robot to move along the predicted trajectory to the sub-target point. The adaptive decision-making and real-time planning involve using an adaptive decision network to select the optimal sequence of boundary vertices from a local map, thereby obtaining the optimal sub-target point. The elevation-aware planning tree then performs accessibility analysis and collision detection based on the optimal sub-target point to generate a path to the sub-target point. The adaptive decision network takes all ordinary vertices and obstacle vertices in the local map and elevation-aware planning tree as input, performs dimensionality reduction mapping on all input data to form a feature map with two-dimensional image features, extracts features from the feature map, evaluates vertices at the map boundary based on the extracted features, selects the optimal boundary vertex sequence, obtains the sub-target point sequence in space through spatial mapping, arranges the sub-target points according to the optimality metric of the sub-target points, and selects the first sub-target point as the optimal sub-target point for the next step. The elevation-aware planning tree includes two structures: a sliding window search tree and a global undirected graph. The sliding window search tree uses the RRT* algorithm to grow in a local map, and its root vertex always moves with the robot's position. During the tree expansion process, the passability of the edges is determined by judging the flatness of the ground, and dangerous areas in the map are marked by obstacle vertices. The global undirected graph is constructed using a probabilistic route graph algorithm without sampling. It is responsible for recording the sub-target points explored by the robot and the paths leading to these sub-target points. Each vertex of the global undirected graph is provided by an adaptive decision network and a sliding window search tree.

6. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-4.

7. A storage medium characterized in that it non-transitory stores computer-readable instructions, wherein, When the non-transitory computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-4 are executed.

Images