A robot navigation path planning method and device based on a multi-level map

Abstract

This application provides a robot navigation path planning method and apparatus based on multi-level maps. The method includes: extracting multi-source perception information from an acquired multi-source perception dataset; mapping the multi-source perception information to a two-dimensional grid space to generate a geometric obstacle layer, a semantic information layer, and a terrain information layer, and constructing multi-level global and local two-dimensional grid maps; determining the cost quantization data of each two-dimensional grid layer based on the multi-source perception information, and generating global and local drivability cost maps; planning a global target path in the global drivability cost map using a target cost search method, and performing local optimization on the global target path based on the local drivability cost map to determine the target navigation path; and controlling the target robot to perform a navigation task according to the motion control commands corresponding to the target navigation path. This method improves the rationality of path planning, navigation stability, and overall safety of the robot in complex environments.

Description

Technical Field

[0001] This application relates to the fields of mobile robot and autonomous driving technology, and in particular to a robot navigation path planning method and apparatus based on multi-level maps. Background Technology

[0002] As the application of mobile robots, service robots, and special-purpose robots in complex indoor and outdoor environments continues to expand, the robot's autonomous navigation capability has become one of its core basic capabilities. Existing robot navigation systems are usually based on two-dimensional geometric grid maps or point cloud maps, which mainly describe the distribution information of static obstacles in the environment and achieve the ability to reach the target point through global path planning and local obstacle avoidance algorithms.

[0003] Current robot navigation path planning methods are mostly based on single-dimensional maps for path search, which has the following technical shortcomings: The map information is limited, only reflecting the distribution of geometric obstacles in the environment. This results in planned paths that can only avoid obstacles, easily leading to unreasonable paths and safety hazards in complex environments. The map update mechanism is also limited; too low an update frequency cannot respond to real-time changes in the dynamic environment, while too high an update frequency leads to excessive computational overhead and map instability, affecting the continuity of path planning. The cost quantification method is imperfect, failing to perform targeted cost modeling for environmental information in different dimensions, resulting in unscientific path planning decisions. Furthermore, the coordination between global and local paths is insufficient, easily leading to problems such as global path unreachability and local obstacle avoidance failures. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a robot navigation path planning method and device based on a multi-level map. By acquiring a multi-source perception dataset and extracting multi-source perception information, a multi-level, multi-dimensional two-dimensional grid map including a geometric obstacle layer, a semantic information layer, and a terrain information layer is constructed. Then, a drivability cost map is generated, a global target path is planned, and local optimization is performed. This realizes the decoupled expression and collaborative use of the robot's multi-dimensional environmental information at different spatial scales, thereby taking into account both global path planning and local motion control requirements, improving the rationality of path planning, navigation stability, and overall safety of the robot in complex environments, and achieving collaborative adaptation between global path stability and local environmental real-time performance.

[0005] This application provides a robot navigation path planning method based on a multi-level map, the method comprising: Obtain the multi-source perception dataset corresponding to the environment in which the target robot is located, and extract multi-source perception information from the multi-source perception dataset; The multi-source sensing information is mapped to a two-dimensional grid space under a unified coordinate system to generate multiple two-dimensional grid layers. Based on the two-dimensional grid layers, multi-level global two-dimensional grid maps and multi-level local two-dimensional grid maps are constructed respectively. The two-dimensional grid layers include a geometric obstacle layer, a semantic information layer, and a terrain information layer. Based on the multi-source sensing information, the cost quantization data corresponding to each of the two-dimensional raster layers is determined, and based on the cost quantization data, the multi-level global two-dimensional raster map, and the multi-level local two-dimensional raster map, a global traversability cost map and a local traversability cost map are generated accordingly. A global target path is planned in the global navigability cost map using the target cost search method, and the global target path is locally optimized based on the local navigability cost map and a preset local cost change threshold to determine the target navigation path. Based on the target navigation path, corresponding motion control commands are generated, and the target robot is controlled to perform navigation tasks according to the motion control commands.

[0006] Furthermore, the step of acquiring the multi-source perception dataset corresponding to the environment in which the target robot is located, and extracting multi-source perception information from the multi-source perception dataset, includes: Obtain a multi-source perception dataset corresponding to the environment in which the target robot is located; wherein, the multi-source perception dataset includes at least visual perception data, semantic recognition data, and terrain perception data; For each type of data in the multi-source sensing dataset, determine whether the time interval step size corresponding to the timestamp of the data type and the preset fusion reference time is less than or equal to the preset time synchronization window. If the time interval step size is less than or equal to the time synchronization window, then this type of data is identified as spatial alignment sensing data. If the time interval step size is larger than the time synchronization window, then according to the pose change relationship of the target robot within the time interval step size, spatiotemporal compensation is performed on this type of data to obtain spatial alignment perception data after compensation for this type of data. Multi-source sensing information is extracted from the spatial alignment sensing data; wherein the multi-source sensing information includes at least geometric obstacle information, semantic category information, and terrain structure information.

[0007] Furthermore, the construction of multi-level global two-dimensional raster maps and multi-level local two-dimensional raster maps based on the two-dimensional raster layer includes: Based on the preset global map raster resolution and local map raster resolution, and the preset global update frequency and local update frequency, a global map and a local map are constructed respectively. A global two-dimensional raster layer is determined in the two-dimensional raster layer, and the global two-dimensional raster layer is merged into the global map. Based on a preset historical memory factor, each raster unit in the merged global map is updated using an incremental fusion update mechanism with historical memory to construct a multi-level global two-dimensional raster map. A local two-dimensional raster layer is determined in the two-dimensional raster layer, and the local two-dimensional raster layer is merged into the local map. Then, each raster unit in the merged local map is updated using an instant update mechanism to construct a multi-level local two-dimensional raster map.

[0008] Furthermore, determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information includes: For the geometric obstacle layer, based on the geometric obstacle information in the multi-source sensing information, the occupancy probability corresponding to each grid cell in the geometric obstacle layer is determined; Based on the occupancy probability and the preset safety threshold, the geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer is determined using a preset monotonically increasing function. The geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer is determined as the cost quantization data corresponding to the geometric obstacle layer.

[0009] Furthermore, the step of determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information further includes: For the semantic information layer, determine the category weight corresponding to each semantic category in the semantic category information of the multi-source perception information; Based on the semantic category information, the semantic category to which each grid cell in the semantic information layer belongs is determined, and the category weight corresponding to the semantic category is determined as the basic semantic cost quantization data corresponding to the grid cell; Based on the time modulation factor corresponding to the grid cell and the basic semantic cost quantization data corresponding to the grid cell, the semantic cost quantization data corresponding to the grid cell is determined; The semantic cost quantization data corresponding to each grid cell in the semantic information layer is determined as the cost quantization data corresponding to the semantic information layer.

[0010] Furthermore, the step of determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information further includes: For the terrain information layer, based on the terrain structure information in the multi-source sensing information, the slope value and surface roughness value corresponding to each grid cell in the terrain information layer are determined respectively; Based on the slope value and the surface roughness value, determine the terrain cost quantization data corresponding to each grid cell in the terrain information layer; The terrain cost quantization data corresponding to each raster cell in the terrain information layer is determined as the cost quantization data corresponding to the terrain information layer.

[0011] Furthermore, the step of generating a global traversability cost map and a local traversability cost map based on the cost quantification data, the multi-level global two-dimensional raster map, and the multi-level local two-dimensional raster map includes: Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level global two-dimensional raster map and the preset cost weight parameters, the global traversability cost quantization data corresponding to each raster unit in the multi-level global two-dimensional raster map is determined; Based on the global traversability cost quantification data and the multi-level global two-dimensional grid map, a global traversability cost map is generated; Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level local two-dimensional raster map and the preset cost weight parameters, the local traversability cost quantization data corresponding to each raster unit in the multi-level local two-dimensional raster map is determined. A local accessibility cost map is generated based on the local accessibility cost quantification data and the multi-level local two-dimensional raster map.

[0012] Furthermore, the step of locally optimizing the global target path based on the local drivability cost map and a preset local cost change threshold to determine the target navigation path includes: Based on the global target path, local optimization is performed in the local drivability cost map using the target cost search method. During the local optimization process, when the change in the local mobility cost quantification data corresponding to the local mobility cost map is greater than a preset local cost change threshold, local replanning is triggered; when the global target path is unreachable or blocked for a long time, global replanning is triggered. In response to the global target path, the local optimization is completed, and the target navigation path is determined.

[0013] This application embodiment also provides a robot navigation path planning device based on a multi-level map, the robot navigation path planning device comprising: The data perception module is used to acquire the multi-source perception dataset corresponding to the environment in which the target robot is located, and to extract multi-source perception information from the multi-source perception dataset. The map building module is used to map the multi-source sensing information to a two-dimensional grid space under a unified coordinate system to generate multiple two-dimensional grid layers, and to construct multi-level global two-dimensional grid maps and multi-level local two-dimensional grid maps based on the two-dimensional grid layers; wherein, the two-dimensional grid layers include a geometric obstacle layer, a semantic information layer and a terrain information layer; The map fusion module is used to determine the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information, and to generate a global traversability cost map and a local traversability cost map based on the cost quantization data, the multi-level global two-dimensional raster map and the multi-level local two-dimensional raster map. The navigation decision module is used to plan a global target path in the global drivability cost map using the target cost search method, and to perform local optimization on the global target path based on the local drivability cost map and a preset local cost change threshold to determine the target navigation path. The execution control module is used to generate corresponding motion control commands based on the target navigation path, and control the target robot to perform navigation tasks according to the motion control commands.

[0014] This application also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the robot navigation path planning method based on a multi-level map as described above are performed.

[0015] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the robot navigation path planning method based on a multi-level map as described above.

[0016] This application provides a robot navigation path planning method and apparatus based on a multi-level map. The method includes: acquiring a multi-source perception dataset corresponding to the environment of the target robot, and extracting multi-source perception information from the multi-source perception dataset; mapping the multi-source perception information to a two-dimensional grid space under a unified coordinate system to generate multiple two-dimensional grid layers, and constructing a multi-level global two-dimensional grid map and a multi-level local two-dimensional grid map based on the two-dimensional grid layers; wherein, the two-dimensional grid layers include a geometric obstacle layer, a semantic information layer, and a terrain information layer; and determining each of the two-dimensional grid layers based on the multi-source perception information. The cost quantization data corresponding to the raster layer is used, and based on the cost quantization data, the multi-level global two-dimensional raster map, and the multi-level local two-dimensional raster map, a global navigability cost map and a local navigability cost map are generated. A global target path is planned in the global navigability cost map using a target cost search method, and the global target path is locally optimized based on the local navigability cost map and a preset local cost change threshold to determine the target navigation path. Based on the target navigation path, corresponding motion control commands are generated, and the target robot is controlled to perform navigation tasks according to the motion control commands.

[0017] Compared with existing path search methods based on single-dimensional maps, this method acquires multi-source perception datasets and extracts multi-source perception information to construct a multi-level, multi-dimensional two-dimensional grid map, including geometric obstacle layers, semantic information layers, and terrain information layers. This generates a drivability cost map, plans the global target path, and performs local optimization. This enables the decoupled expression and collaborative use of the robot's multi-dimensional environmental information at different spatial scales, thus balancing global path planning and local motion control requirements. It improves the rationality of path planning, navigation stability, and overall safety of the robot in complex environments, achieving a coordinated adaptation between global path stability and local environmental real-time performance.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1A flowchart illustrating a robot navigation path planning method based on a multi-level map, provided as an embodiment of this application; Figure 2 This is a schematic diagram illustrating a method for generating a accessibility cost map based on a two-dimensional raster layer, as provided in an embodiment of this application. Figure 3 This is a schematic diagram of a multi-source sensing information mapping accessibility cost map provided in an embodiment of this application; Figure 4 A schematic diagram of the structure of a robot navigation path planning device based on a multi-level map provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.

[0022] Research has found that existing robot navigation systems are usually based on two-dimensional geometric grid maps or point cloud maps, which mainly describe the distribution information of static obstacles in the environment and achieve the ability to reach the target point through global path planning and local obstacle avoidance algorithms.

[0023] However, current robot navigation path planning methods are mostly based on single-dimensional maps for path search, which has the following technical shortcomings: The map information is limited, only reflecting the distribution of geometric obstacles in the environment, resulting in planned paths that can only avoid obstacles. In complex environments, this can easily lead to unreasonable paths and safety hazards. The map update mechanism is also limited; too low an update frequency cannot respond to real-time changes in the dynamic environment, while too high an update frequency leads to excessive computational overhead and map instability, affecting the continuity of path planning. The cost quantification method is imperfect, failing to perform targeted cost modeling for environmental information in different dimensions, resulting in unscientific path planning decisions. Furthermore, the coordination between global and local paths is insufficient, easily leading to problems such as global path unreachability and local obstacle avoidance failures.

[0024] Furthermore, the heterogeneous information corresponding to the environmental data acquired by the robot often exists in an independent form, lacking a unified map representation and fusion mechanism, making it difficult to directly serve the navigation decision-making and path planning process, resulting in high system complexity, poor scalability, and high engineering implementation costs.

[0025] Based on this, this application provides a robot navigation path planning method based on a multi-level map. By acquiring a multi-source perception dataset and extracting multi-source perception information, a multi-level, multi-dimensional two-dimensional grid map including a geometric obstacle layer, a semantic information layer, and a terrain information layer is constructed. Then, a drivability cost map is generated, a global target path is planned, and local optimization is performed. This enables the decoupled expression and collaborative use of the robot's multi-dimensional environmental information at different spatial scales, thereby taking into account both global path planning and local motion control requirements. This improves the rationality of the robot's path planning, navigation stability, and overall safety in complex environments, achieving a collaborative adaptation between global path stability and local environmental real-time performance.

[0026] Please see Figure 1 , Figure 1 This is a flowchart illustrating a robot navigation path planning method based on a multi-level map, provided as an embodiment of this application. Figure 1 As shown in the embodiments of this application, the robot navigation path planning method based on multi-level maps is generally applicable to scenarios where indoor and outdoor mobile robots, unmanned delivery vehicles, service robots, and special operation robots need to navigate autonomously in complex and dynamic environments. The method includes: S101. Obtain the multi-source perception dataset corresponding to the environment where the target robot is located, and extract multi-source perception information from the multi-source perception dataset.

[0027] In this embodiment of the application, the multi-source perception dataset includes, but is not limited to, visual perception data, semantic recognition data, and terrain perception data; the multi-source perception information includes, but is not limited to, geometric obstacle information, semantic category information, and terrain structure information.

[0028] In this step, various types of raw perception data of the environment are collected by the multiple sensors carried by the target robot to form a multi-source perception dataset. The dataset is then preprocessed with time synchronization, spatiotemporal compensation, etc., to extract key environmental information for subsequent map construction and cost calculation.

[0029] In this way, by acquiring comprehensive and accurate environmental information required for robot navigation, problems such as insufficient perception dimensions of a single sensor, data asynchrony leading to subsequent map construction deviations and cost calculation errors can be solved. This provides reliable data support for multi-level map construction and path planning, enables synchronous collection and standardized processing of multi-dimensional environmental information, ensures that the extracted multi-source perception information has spatiotemporal consistency, covers the three core dimensions of geometry, semantics, and terrain, and eliminates perception blind spots and data misalignment.

[0030] In one possible implementation of this application, step S101 may include: S1011. Obtain the multi-source perception dataset corresponding to the environment in which the target robot is located.

[0031] In this step, the target robot uses multiple sensors to simultaneously collect raw perception data of its environment, forming a multi-source perception dataset.

[0032] Visual perception data is collected by visual sensors (such as cameras and depth cameras) on the robot, including environmental images and depth information, and is used to identify obstacles and target outlines in the environment, providing a basis for extracting geometric obstacle information. Semantic recognition data is collected by semantic recognition sensors (such as visual semantic sensors and LiDAR) on the robot, or obtained by semantic segmentation processing of visual perception data, and is used to identify the semantic categories of various targets or areas in the environment (such as pedestrians, vehicles, passable ground, restricted areas, etc.). Terrain perception data is collected by terrain perception sensors (such as LiDAR, ultrasonic sensors, and inertial measurement units) on the robot, and is used to obtain terrain parameters of the environmental surface (such as slope, surface roughness, and undulation).

[0033] For example, suppose an outdoor inspection robot is selected as the target robot. This robot is equipped with a high-definition camera (to collect visual perception data), a laser semantic radar (to collect semantic recognition data and some geometric perception data), an inertial measurement unit (to collect terrain perception data), and an ultrasonic sensor (to assist in collecting terrain data). In an outdoor industrial park scenario, image data (visual perception data) of the park, semantic category data of pedestrians, vehicles, roads, walls, etc. (semantic recognition data), and data such as the slope (0~15°) and surface roughness (0.1~0.5mm) of the park roads (terrain perception data) are collected simultaneously to form a multi-source perception dataset. The collection frequencies of each type of data are as follows: visual perception data 30 frames / second, semantic recognition data 10 frames / second, and terrain perception data 50Hz.

[0034] S1012. For each type of data in the multi-source sensing dataset, determine whether the time interval step size corresponding to the timestamp of the data type and the preset fusion reference time is less than or equal to the preset time synchronization window.

[0035] In the embodiments of this application, since different sensors have different acquisition frequencies and response speeds, the timestamps (acquisition times) of various types of sensing data are different. If used directly, it will lead to data spatiotemporal misalignment. Therefore, time synchronization judgment is required.

[0036] Wherein: the timestamp is the collection time recorded when each type of sensing data is collected, used to identify the time attribute of the data; the fusion reference time is a preset reference time used to unify various types of sensing data, and all sensing data must be aligned to this reference time; the time interval step size is the absolute value of the difference between the timestamp of a certain type of sensing data and the fusion reference time; the time synchronization window is a preset time threshold used to determine whether the data can be directly aligned. When the time interval step size is less than or equal to this threshold, the time deviation of the data can be ignored and can be directly used for subsequent spatial alignment; when it is greater than this threshold, spatiotemporal compensation is required.

[0037] In this embodiment of the application, the expression is: To determine whether the time interval step size corresponding to the timestamp of this type of data and the preset fusion reference time is less than or equal to the preset time synchronization window; where, Indicates the first The timestamp corresponding to the data type This indicates the preset fusion reference time. This indicates the preset time synchronization window.

[0038] S1013. If the time interval step size is less than or equal to the time synchronization window, then this type of data is identified as spatial alignment sensing data.

[0039] In this step, for the sensing data whose time interval step size is within the time synchronization window, the time deviation is small and will not affect the accuracy of subsequent spatial alignment. Therefore, this type of data is directly identified as spatial alignment sensing data, without the need for additional time compensation. It only needs to be mapped to a unified coordinate system later.

[0040] S1014. If the time interval step size is larger than the time synchronization window, then according to the pose change relationship of the target robot within the time interval step size, spatiotemporal compensation is performed on this type of data to obtain spatial alignment perception data after compensation for this type of data.

[0041] In this step, the perception data whose time interval step size exceeds the time synchronization window has a large time deviation. If used directly, it will lead to spatial misalignment. Therefore, it is necessary to perform spatiotemporal compensation on the data according to the robot's pose change within the time interval and align it to the spatial position corresponding to the fusion reference time.

[0042] Here, the pose change relationship is: the robot's step size in the time interval. The robot's position and attitude changes are acquired through devices such as odometry and inertial measurement units, and then processed using a pose transformation matrix. This indicates that the matrix can... Real-time collected data coordinates are converted into The coordinates of the moment.

[0043] In this embodiment, the spatiotemporal compensation calculation is performed by transforming the data using a pose transformation matrix, and the compensation formula is shown below.

[0044] .

[0045] in, for Raw sensory data at any given moment; To compensate and align to Spatial alignment sensing data at any given time; This is the pose transformation matrix.

[0046] S1015. Extract multi-source sensing information from the spatial alignment sensing data.

[0047] In this step, information is extracted from all spatially aligned sensing data (both directly determined and compensated) to obtain three types of core multi-source sensing information.

[0048] Among them, geometric obstacle information is extracted from visual perception data and laser perception data (semantic radar assisted), including the spatial coordinates, outline, and occupancy range of obstacles, which is used for the construction of subsequent geometric obstacle layers; semantic category information is extracted from semantic recognition data, including the semantic categories of various targets or areas in the environment (such as passable ground, stairs, pedestrians, vehicles, restricted areas, static obstacles, etc.), which is used for the construction of subsequent semantic information layers; terrain structure information is extracted from terrain perception data, including parameters such as slope, surface roughness, and undulation of the ground surface, which is used for the construction of subsequent terrain information layers.

[0049] S102. The multi-source sensing information is mapped to a two-dimensional grid space under a unified coordinate system to generate multiple two-dimensional grid layers. Based on the two-dimensional grid layers, a multi-level global two-dimensional grid map and a multi-level local two-dimensional grid map are constructed respectively.

[0050] The two-dimensional raster layer includes a geometric obstacle layer, a semantic information layer, and a terrain information layer.

[0051] Here, a two-dimensional raster layer maps single-dimensional environmental information (geometry, semantics, terrain) into a raster set. Each layer contains only one type of environmental information, which facilitates subsequent independent updates and cost calculations.

[0052] The geometric obstacle layer describes the spatial distribution of obstacles in the environment. Each grid cell corresponds to an "occupied" (obstacles exist) or "idle" (no obstacles) state, which is the core basis for avoiding obstacles in path planning. The semantic information layer describes the semantic categories of various targets or areas in the environment (such as passable ground, stairs, pedestrians, vehicles, restricted areas, etc.). Each grid cell corresponds to a specific semantic category, providing a basis for semantic risk judgment in path planning. The terrain information layer describes the terrain attributes of the environment's surface (such as slope, surface roughness, etc.). Each grid cell corresponds to a specific terrain parameter, providing a basis for terrain adaptability judgment in path planning.

[0053] In this step, the extracted multi-source perception information is first uniformly mapped to a preset two-dimensional grid space (that is, the continuous environmental information is discretized into grid units through grid processing) to generate three types of two-dimensional grid layers corresponding to geometric, semantic and terrain information respectively; then, based on these three types of layers, a global two-dimensional grid map for large-scale navigation and a local two-dimensional grid map for local fine obstacle avoidance are constructed respectively, forming a multi-level map system.

[0054] In the embodiments of this application, a unified coordinate system refers to a preset, fixed spatial coordinate system used to unify the spatial position of all perceived information and maps, and to avoid spatial misalignment caused by inconsistencies in coordinate systems. Typically, a robot base coordinate system or a world coordinate system is used. The two-dimensional grid space divides the two-dimensional environment in which the robot is located into several uniformly sized square grid cells. Each grid cell corresponds to a specific spatial position. The environmental information is described by the attributes of the grid cells (such as occupancy status, semantic category, and terrain parameters), and it is the basic carrier for map construction.

[0055] Here, the multi-level global two-dimensional grid map is constructed based on global environmental information, covering the entire area of ​​robot navigation. It has a low grid resolution and a slow update frequency, and is used to describe the overall structure of a large-scale environment to support global path planning. The multi-level local two-dimensional grid map is constructed based on environmental information in a small area around the robot, covering only a certain area around the robot. It has a high grid resolution and a fast update frequency, and is used to describe the detailed changes in the local environment to support local path optimization and real-time obstacle avoidance.

[0056] In the embodiments of this application, the representation forms of multi-level global two-dimensional raster maps and multi-level local two-dimensional raster maps are as follows.

[0057] ; .

[0058] in, Represents a multi-level global two-dimensional raster map; , , These correspond to the global two-dimensional raster layers in a multi-level global two-dimensional raster map (i.e., the global geometric obstacle layer, the global semantic information layer, and the global terrain information layer). Represents a multi-level local two-dimensional raster map; , , These correspond to the local two-dimensional raster layers in a multi-level local two-dimensional raster map (i.e., local geometric obstacle layer, local semantic information layer, and local terrain information layer).

[0059] This approach enables layered and scaled representation of environmental information, transforming heterogeneous multi-source sensing information into a structured, computable raster map. It distinguishes the functional positioning of global and local maps, addressing the issues of traditional single-map information representation being limited and unable to balance global stability with local real-time performance. The resulting multi-level map system can reflect the large-scale environmental structure through the global map, providing a stable basis for global path planning, while also capturing local environmental details and dynamic changes through the local map, providing precise support for local optimization. The independent design of the three types of layers facilitates subsequent independent updates and cost calculations, reducing system complexity.

[0060] In one possible implementation of this application, in specific implementation, step S102, based on the two-dimensional raster layer, may include the following steps: S1021. Construct a global map and a local map based on preset global map raster resolution and local map raster resolution, and preset global update frequency and local update frequency.

[0061] In this step, the core parameters (resolution, update frequency) of the global map and the local map are preset first. Based on these parameters, the basic framework (i.e., the set of blank grids) of the global map and the local map is built, which provides a carrier for subsequent layer fusion and updates.

[0062] Among them, grid resolution refers to the actual spatial size corresponding to each grid cell. The higher the resolution, the smaller the grid cell and the finer the map details. In the embodiments of this application, the global map grid resolution is greater than the local map grid resolution, so that the local map has a more refined spatial expression capability (the global map focuses on large-scale structure, while the local map focuses on details). The update frequency refers to the frequency of grid cell state updates in the map. The global map update frequency is much lower than the local map update frequency to achieve a rapid response to dynamic environmental changes (the global map focuses on stability, while the local map focuses on real-time performance).

[0063] S1022. Determine a global two-dimensional raster layer in the two-dimensional raster layer, merge the global two-dimensional raster layer into the global map, and update each raster unit in the merged global map based on a preset historical memory factor using an incremental fusion update mechanism with historical memory, so as to construct a multi-level global two-dimensional raster map.

[0064] In this step, firstly, grid data covering the entire global area is selected from the three types of generated two-dimensional grid layers and used as global two-dimensional grid layers; then, these three types of global two-dimensional grid layers are merged into the preset global map frame; finally, an incremental fusion update mechanism with historical memory is used to update each grid unit of the global map, ultimately forming a multi-level global two-dimensional grid map containing the three types of layers.

[0065] Among them, the global two-dimensional grid layer is grid data extracted from the original two-dimensional grid layer that covers the entire area of ​​robot navigation and has the same spatial range as the global map.

[0066] Here, the historical memory factor is a preset weight parameter (0<λ<1) used to balance historical map information and current observation information. The larger λ is, the more historical information is valued and the more stable the map is; the smaller λ is, the more current observation is valued and the more flexible the map is.

[0067] In this embodiment of the application, the incremental fusion update mechanism with historical memory is the core update method of the global map. Instead of directly overwriting the historical map with the current single observation data, it updates the raster cell state through weighted fusion. The update formula of the incremental fusion update mechanism with historical memory is as follows.

[0068] .

[0069] in, For each grid cell in the global map at time t, ... The final state (including geometric, semantic, and terrain attributes); For the raster cell at time t-1 (previous update cycle) The historical state; For the grid cell at time t The current observation status; It is a factor of historical memory.

[0070] Here, incremental fusion only updates the calculations for newly added sensing areas or raster cells whose states have changed. Unchanged raster cells directly inherit the historical state, reducing computational overhead.

[0071] S1023. Determine a local two-dimensional raster layer in the two-dimensional raster layer, merge the local two-dimensional raster layer into the local map, and use an instant update mechanism to update each raster unit in the merged local map to construct a multi-level local two-dimensional raster map.

[0072] In this step, firstly, grid data covering a small area around the robot is selected from the three types of generated two-dimensional grid layers and used as local two-dimensional grid layers; then, these three types of local two-dimensional grid layers are merged into a preset local map frame; finally, an instant update mechanism is used to update each grid unit of the local map, ultimately forming a multi-level local two-dimensional grid map containing the three types of layers.

[0073] Among them, the local two-dimensional raster layer is a small-scale raster data centered on the robot extracted from the original two-dimensional raster layer, which is consistent with the spatial range of the local map.

[0074] In this embodiment of the application, the real-time update mechanism is the core update method of the local map. It directly uses the latest observation data to overwrite the state of the grid cells at the previous moment, without weighted fusion of historical information, thereby enhancing the real-time adaptability to dynamic targets. The update formula of the real-time update mechanism is as follows.

[0075] .

[0076] in, For the local map at time t, the raster unit The state is directly determined by the current observation data at time t, replacing the state at time t-1.

[0077] Here, a full update means replacing all grid cells in the local map each time it is updated, ensuring that the local map can reflect the changes in environmental details around the robot in real time.

[0078] S103. Based on the multi-source sensing information, determine the cost quantization data corresponding to each of the two-dimensional raster layers, and generate a global traversability cost map and a local traversability cost map based on the cost quantization data, the multi-level global two-dimensional raster map, and the multi-level local two-dimensional raster map.

[0079] In this step, the qualitative / quantitative environmental information (such as obstacle occupancy status, semantic category, and terrain parameters) of each of the three types of generated two-dimensional raster layers is first converted into quantifiable cost data (i.e., cost quantification data) by combining the extracted multi-source perception information. Then, the cost quantification data of each type of layer is fused together to correspond to the global map and the local map, respectively, to generate a global accessibility cost map and a local accessibility cost map that can reflect the difficulty and risk of passage in the area.

[0080] In this embodiment of the application, the cost quantification data is the conversion of the environmental attributes (such as obstacle occupancy probability, semantic category, terrain parameters) of each grid cell in the two-dimensional grid layer into a numerical value with a unified dimension. This numerical value is used to characterize the difficulty, risk or cost of the robot passing through the grid cell. The larger the value, the higher the difficulty / risk of passage.

[0081] Here, the traversability cost map uses a grid as its carrier, and the value of each grid cell is the comprehensive traversal cost of that cell. It is the core basis for searching the optimal path in path planning. The lower the comprehensive cost, the better the path (safer and more efficient). It is divided into two categories: global traversability cost map and local traversability cost map, which correspond to the scale of the global map and the local map, respectively.

[0082] In this way, heterogeneous environmental information (geometric, semantic, and topographical) is transformed into a unified and computable comprehensive cost, providing a scientific basis for path planning. This solves the problem that traditional path planning only considers geometric obstacles and ignores semantic risks and terrain adaptability, thereby improving the rationality and safety of path planning.

[0083] Here, the generated traversability cost map can comprehensively reflect the traversal difficulty and risk of each grid cell, taking into account both the occlusion constraints of geometric obstacles and semantic risks (such as restricted areas and dynamic targets) and terrain adaptability (such as steep slopes and rough ground), providing accurate cost support for subsequent global path planning and local optimization.

[0084] In this embodiment of the application, in terms of the cost setting of each two-dimensional raster layer, a multi-dimensional comprehensive cost model is constructed to achieve unified modeling of geometric security, semantic risk and terrain adaptability.

[0085] In one possible implementation of this application, in specific implementation, for a geometric obstacle layer, the step S103 of determining the cost quantization data corresponding to each of the two-dimensional grid layers based on the multi-source sensing information may include: S10311. For the geometric obstacle layer, based on the geometric obstacle information in the multi-source sensing information, determine the occupancy probability corresponding to each grid cell in the geometric obstacle layer.

[0086] In this embodiment, the core of cost quantization for the geometric obstacle layer is the occupancy probability of the grid cell, which is used to characterize the possibility of an obstacle existing within the grid cell.

[0087] In this step, based on the extracted geometric obstacle information (obstacle spatial coordinates, outline, occupancy range, etc.), the occupancy probability of each grid cell is determined through a probability calculation model.

[0088] The occupancy probability ranges from [0,1], where 0 represents absolutely free (no obstacles) and 1 represents absolutely occupied (with obstacles). The larger the value, the higher the probability of the existence of obstacles.

[0089] S10312. Based on the occupancy probability and the preset safety threshold, the geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer is determined using a preset monotonically increasing function.

[0090] In this step, the occupancy probability of grid cells is converted into a quantifiable geometric cost. A monotonically increasing function is used to ensure that the higher the occupancy probability, the greater the geometric cost (the higher the difficulty of passage). At the same time, the cost is constrained by a preset safety threshold to avoid unreasonable costs due to small occupancy probabilities.

[0091] Among them, the safety threshold is a preset occupancy probability threshold used to distinguish between "safe grids" and "dangerous grids". When the occupancy probability is below the threshold, the cost increases slowly; when it is above the threshold, the cost rises rapidly and non-linearly; when it exceeds the danger threshold, the cost tends to the upper limit; the monotonically increasing function is a function used to map the occupancy probability to geometric cost, and this function satisfies the monotonically increasing characteristic.

[0092] In this embodiment of the application, the geometric cost quantization data is determined by the following monotonically increasing function.

[0093] .

[0094] in, This represents the geometric cost quantization data; It is a monotonically increasing function; This represents the probability of occupancy.

[0095] S10313. Determine the geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer as the cost quantization data corresponding to the geometric obstacle layer.

[0096] In this embodiment of the application, the cost quantization data of the geometric obstacle layer is the collection of geometric cost quantization data of all grid cells in the layer. Each grid cell corresponds to a unique geometric cost value, which will be subsequently fused with semantic and terrain cost data to generate a drivability cost map.

[0097] In one possible implementation of this application, in specific implementation, for the semantic information layer, the step of determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information in step S103 further includes: S10321. For the semantic information layer, determine the category weight corresponding to each semantic category in the semantic category information of the multi-source perception information.

[0098] In this embodiment, the core of the cost quantification of the semantic information layer is the semantic category weight. Different semantic categories have different risks and difficulties in passage. Therefore, a corresponding category weight is preset for each semantic category. The larger the weight value, the higher the risk / difficulty of passage for that semantic category.

[0099] Here, the semantic category set is preset to "{accessible ground, stairs, slopes, pedestrians, vehicles, static obstacles, restricted areas}", and the weight of each category is set according to actual navigation needs and can be adjusted adaptively.

[0100] For example, a set of semantic categories S and corresponding category weights are preset. The following are accessible ground surfaces: =0.1, low risk), stairs ( =0.8, high risk, robots have difficulty passing through), slope ( =0.3~0.6, adjusted according to slope, medium risk), pedestrians ( =0.5, dynamic risk, requires deceleration and avoidance), vehicle ( =0.7, dynamic risk, needs to be avoided), static obstacles ( =0.9, high risk, impassable), restricted area ( =1.0, extremely high risk, passage prohibited); this weight setting is adapted to the navigation needs of outdoor inspection robots, prioritizing the avoidance of high-risk semantic categories.

[0101] In the embodiments of this application, different semantic categories correspond to different risk levels. For example, passable ground has the lowest weight; slopes are given medium weight according to their slope level; stairs can be set to extremely high cost in wheeled robot application scenarios; restricted areas can be directly set to an impassable state. Through the semantic weighting mechanism, the planning results have environmental understanding capabilities.

[0102] S10322. Based on the semantic category information, determine the semantic category to which each grid cell in the semantic information layer belongs, and determine the category weight corresponding to the semantic category as the basic semantic cost quantization data corresponding to the grid cell.

[0103] In this step, the semantic category corresponding to each grid cell in the semantic information layer is determined by combining the extracted semantic category information (e.g., the semantic category of a certain grid cell is "accessible ground"). Then, the preset category weight of the semantic category is used as the basic semantic cost quantification data of the grid cell.

[0104] In this embodiment of the application, the basic semantic cost quantization number is determined by the following expression.

[0105] .

[0106] in, Indicates the first The category weights corresponding to each semantic category; Represents grid cells The corresponding basic semantic cost quantification data.

[0107] S10323. Based on the time modulation factor corresponding to the grid cell and the basic semantic cost quantization data corresponding to the grid cell, determine the semantic cost quantization data corresponding to the grid cell.

[0108] In this step, for dynamic semantic categories (such as pedestrians and vehicles), their passage risk changes over time (such as changes in pedestrian movement speed and distance from the robot). Therefore, a time modulation factor is introduced to modulate the basic semantic cost to obtain the final semantic cost quantification data.

[0109] The time modulation factor is used to characterize the dynamic risk of dynamic semantic categories at time t, and is related to the target speed, prediction uncertainty, and robot distance: the faster the target speed, the closer the distance to the robot, and the higher the prediction uncertainty, the larger the time modulation factor; the time modulation factor of static semantic categories (such as passable ground, static obstacles, and restricted areas) is 1, that is, no modulation.

[0110] Here, the time modulation factor is a function of the target speed, prediction uncertainty, and distance to the robot. When the dynamic target speed increases or the distance decreases, the cost of the corresponding area automatically increases, thereby guiding the planned path to avoid risks in advance and achieving predictive risk control.

[0111] In this embodiment of the application, the semantic cost calculation formula is as follows.

[0112] .

[0113] in, for Time grid unit Semantic cost quantification data, Indicates the first The category weights (i.e., grid cells) corresponding to each semantic category. (corresponding basic semantic cost quantization data); This is the time modulation factor.

[0114] S10324. The semantic cost quantization data corresponding to each grid cell in the semantic information layer is determined as the cost quantization data corresponding to the semantic information layer.

[0115] In the embodiments of this application, the cost quantization data of the semantic information layer is the set of semantic cost quantization data of all grid cells in the layer (static semantic categories without modulation, dynamic semantic categories after modulation). Each grid cell corresponds to a unique semantic cost value, which, together with geometric cost data and terrain cost data, is used to generate a navigability cost map.

[0116] In one possible implementation of this application, in specific implementation, for the terrain information layer, the step S103 of determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information further includes: S10331. For the terrain information layer, based on the terrain structure information in the multi-source sensing information, determine the slope value and surface roughness value corresponding to each grid cell in the terrain information layer.

[0117] In this embodiment, the core of cost quantization of the terrain information layer is the terrain parameters (slope, surface roughness). Based on the extracted terrain structure information (slope, surface roughness, undulation, etc.), the actual terrain parameters corresponding to each grid cell are mapped to the grid cells of the terrain information layer to obtain the slope value and surface roughness value of each grid cell.

[0118] Among them, the slope value is the angle between the ground surface corresponding to the grid cell and the horizontal plane, and the value range is [0°, 90°]. The larger the angle, the steeper the terrain and the higher the difficulty of passage. The surface roughness value is the roughness of the ground surface corresponding to the grid cell, and the value range is [0mm, 1mm]. The larger the value, the rougher the ground surface, the greater the bumps and resistance when the robot moves, and the higher the difficulty of passage.

[0119] S10332. Based on the slope value and the surface roughness value, determine the terrain cost quantization data corresponding to each grid cell in the terrain information layer.

[0120] In this step, the slope value and surface roughness value are converted into quantifiable terrain costs. A weighted summation method is used, combined with preset weight coefficients, to calculate the quantified terrain cost data for each grid cell.

[0121] In this embodiment of the application, the terrain cost quantification data is determined by the following formula.

[0122] .

[0123] in, Represents grid cells Quantitative data on terrain costs; Represents grid cells The slope value; Represents grid cells Surface roughness value; and This represents the preset weighting coefficients used to adjust the degree of influence of slope and surface roughness on terrain cost. These can be set according to the robot's terrain adaptability (e.g., if the robot is more sensitive to slope, the weighting coefficients can be adjusted). =0.02, =0.01).

[0124] Here, the squared term is designed to achieve the effect that "the larger the terrain parameter, the faster the cost increases", avoiding excessive cost fluctuations caused by small parameter changes; when the slope value exceeds the preset safety threshold, the grid cell is marked as impassable, and the terrain cost tends to the upper limit.

[0125] S10333, The terrain cost quantization data corresponding to each grid cell in the terrain information layer is determined as the cost quantization data corresponding to the terrain information layer.

[0126] In this embodiment of the application, the cost quantization data of the terrain information layer is the collection of terrain cost quantization data of all grid cells in the layer. Each grid cell corresponds to a unique terrain cost value, which intuitively reflects the terrain accessibility difficulty of the grid cell. It will be subsequently integrated with geometric cost data and semantic cost data to jointly support the generation of global and local accessibility cost maps.

[0127] In one possible implementation of this application, in specific implementation, the step S103 of generating a global traversability cost map and a local traversability cost map based on the cost quantification data, the multi-level global two-dimensional grid map, and the multi-level local two-dimensional grid map may include: S1034. Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level global two-dimensional raster map and the preset cost weight parameters, determine the global traversability cost quantization data corresponding to each raster unit in the multi-level global two-dimensional raster map.

[0128] S1035. Based on the global traversability cost quantification data and the multi-level global two-dimensional grid map, generate a global traversability cost map.

[0129] S1036. Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level local two-dimensional raster map and the preset cost weight parameters, determine the local accessibility cost quantization data corresponding to each raster unit in the multi-level local two-dimensional raster map.

[0130] S1037. Based on the local accessibility cost quantification data and the multi-level local two-dimensional raster map, generate a local accessibility cost map.

[0131] For example, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating a method for generating a accessibility cost map based on a two-dimensional raster layer, as provided in an embodiment of this application. Figure 2 As shown, the geometric obstacle layer provides basic space occupancy constraints to ensure physical obstacle avoidance safety; the semantic information layer provides environmental understanding constraints to achieve functional area differentiation and risk classification management; and the terrain information layer provides motion accessibility constraints to characterize the impact of slope and surface characteristics on motion stability. After being coupled by weights, the three layers form a unique comprehensive cost value on each grid cell, thereby generating a global accessibility cost map or a local accessibility cost map.

[0132] In this embodiment of the application, global mobility cost quantization data and local mobility cost quantization data are determined by the following expression of the comprehensive cost function.

[0133] .

[0134] in, Represents each grid cell in a multi-level global two-dimensional raster map. The corresponding global accessibility cost quantification data or each grid cell in a multi-level local two-dimensional raster map The corresponding local accessibility cost quantification data; , , These represent the cost quantification data; , , This represents the preset cost weight parameters corresponding to geometry, semantics, and terrain, respectively.

[0135] In this embodiment of the application, a global navigability cost map and a local navigability cost map are generated using the following mapping expression.

[0136] .

[0137] in, This represents a global or local traversability cost map. This represents the quantification data of global or local accessibility costs.

[0138] here, A scalar field defined on a uniform grid space, each of its grid cells The value reflects the overall difficulty or risk level of passage at that location, that is, the cost. The larger the value, the higher the risk or cost of passage in the area; contemporary value When the preset upper limit threshold is exceeded, the corresponding grid can be marked as an impassable area.

[0139] Furthermore, in another embodiment of this application, in order to ensure the consistency of different sub-cost dimensions, the various cost quantification data are normalized before fusion so that their numerical range is mapped to a unified interval, such as [0,1] or [0,C_max], thereby avoiding a single layer from dominating the overall decision due to excessively large numerical scale. The normalized cost quantification data are then weighted and fused through weight coefficients to generate a continuous and smooth comprehensive cost distribution.

[0140] Here, the global and local navigability cost maps are not only used for cost accumulation calculation in the path search algorithm, but also for local obstacle avoidance control, speed adjustment, and risk assessment. For example, in the local planning stage, the robot speed can be modulated according to the average cost value of the grids near the current path, so that the robot automatically reduces its speed in high-risk areas and maintains high operating efficiency in low-risk areas. Thus, the global and local navigability cost maps not only undertake path selection functions, but also safety control and behavior adjustment functions.

[0141] For example, please refer to Figure 3 , Figure 3 This is a schematic diagram illustrating a multi-source sensing information mapping accessibility cost map provided in an embodiment of this application. Figure 3As shown, by fusing geometric obstacle information, semantic category information, and terrain structure information under a unified cost model, a multi-dimensional information-coupled global drivability cost map and a local drivability cost map are constructed. This map maintains spatial consistency, has semantic interpretability, and satisfies motion constraints at the physical level, thus providing a unified, stable, and scalable decision-making basis for subsequent global and local planning.

[0142] S104. Using the target cost search method, a global target path is planned in the global navigability cost map, and based on the local navigability cost map and a preset local cost change threshold, the global target path is locally optimized to determine the target navigation path.

[0143] In this step, the target cost search method is first used to search for the path with the lowest overall cost in the global drivability cost map, which is used as the global target path (i.e. the path framework for large-scale navigation). Then, combined with the local drivability cost map, the global target path is finely optimized, while the changes in local cost are monitored in real time. When the change exceeds a preset threshold or the global path becomes unreachable, the corresponding replanning is triggered, and finally the target navigation path for the robot to actually travel is determined.

[0144] In the embodiments of this application, the target cost search method refers to an algorithm for optimal path search based on a cost map. The core is to find a path from the starting point to the destination with the lowest overall travel cost. An appropriate algorithm can be selected according to the actual scenario.

[0145] Here, the global target path is a path planned based on the global drivability cost map, covering the entire starting point to the end point of the robot navigation. It is mainly used to determine the navigation direction and path framework in a large range, without considering the detailed changes in the local environment. The target navigation path is the actual driving path of the robot that is finally determined after local optimization and replanning. It conforms to the global navigation framework and adapts to the details and dynamic changes of the local environment. It is the optimal path that takes into account both global stability and local real-time performance.

[0146] Here, the local cost change threshold is a preset value used to determine whether the local environment has changed significantly. When the change in the cost quantification data of the local accessibility cost map exceeds this threshold, it indicates that the local environment has changed dynamically (such as the movement of obstacles or the addition of temporary obstacles), and local replanning needs to be triggered.

[0147] This enables collaborative planning of global and local paths, balancing the stability of global paths with the real-time performance of local paths. It solves the problems of unreachable global paths and untimely obstacle avoidance in traditional path planning, ensuring the feasibility and safety of the path.

[0148] Here, the planned target navigation path can not only meet the needs of large-scale navigation, but also accurately adapt to the dynamic changes of the local environment. Through the replanning mechanism, it can respond to environmental changes in a timely manner (such as obstacle movement or path blockage), avoid navigation failure, and improve the robustness of robot navigation.

[0149] In this embodiment of the application, the expression for planning a global target path in the global drivability cost map using the target cost search method is as follows.

[0150] .

[0151] in, Indicates the global target path; This represents the quantified global accessibility cost data corresponding to each grid cell in the global accessibility cost map.

[0152] In one possible implementation of this application, in specific implementation, the step S104 of performing local optimization on the global target path based on the local navigability cost map and a preset local cost change threshold to determine the target navigation path may include: S1041. Based on the global target path, perform local optimization in the local drivability cost map using the target cost search method.

[0153] In this embodiment of the application, the expression for local optimization using the target cost search method in the local drivability cost map is as follows.

[0154] .

[0155] in, Indicates the target navigation path; This represents the quantified local accessibility cost data corresponding to each grid cell in the local accessibility cost map.

[0156] S1042. During the local optimization process, when the change in the local accessibility cost quantification data corresponding to the local accessibility cost map is greater than the preset local cost change threshold, local replanning is triggered; when the global target path is unreachable or blocked for a long time, global replanning is triggered.

[0157] In this embodiment of the application, during the local optimization process, it is necessary to monitor the cost changes of the local drivability cost map in real time to avoid safety hazards in the optimized path due to sudden changes in the local environment; at the same time, it is necessary to monitor the reachability of the global target path to ensure the effectiveness of the global navigation framework.

[0158] Specifically, local replanning is triggered by calculating in real time the change in cost quantification data of the corresponding locally optimized road segment in the local accessibility cost map between two adjacent update cycles. When the change exceeds the local cost change threshold, it indicates that the local environment has undergone significant dynamic changes (such as sudden pedestrians, temporary obstacle movement, or sudden terrain changes). The original locally optimized path is no longer suitable for the current environment, and local replanning needs to be triggered immediately to search for the local optimal sub-path again to ensure the timeliness of local obstacle avoidance.

[0159] Furthermore, global replanning is triggered: the accessibility of the global target path is judged in real time. When the global traversability cost of multiple consecutive grid cells on the global target path is greater than the preset impassable threshold, or the global path is blocked for a long time (such as construction closure, the addition of permanent obstacles, and the blocking time exceeding the preset duration), it means that the original global target path is no longer traversable and global replanning needs to be triggered. Based on the updated global traversability cost map, the global target path is replanned to ensure the smooth completion of the navigation task.

[0160] S1043. In response to the global target path, the local optimization is completed, and the target navigation path is determined.

[0161] In this step, once the local optimization of the global target path is completed and no local replanning has been triggered, or after local replanning and global replanning have been completed, the locally optimized path segment is seamlessly connected with the unoptimized global target path segment to form a complete target navigation path.

[0162] In this embodiment, the target navigation path retains the overall framework and large-scale navigation rationality of the global target path, while incorporating real-time details and dynamic changes of the local environment. This achieves a synergy between global stability and local real-time performance, making it a directly drivable path for the robot to ultimately perform navigation tasks. At the same time, the target navigation path must meet the requirements of path continuity, traffic safety, and efficiency to ensure that the robot can smoothly and safely complete the navigation task from the starting point to the destination along the path.

[0163] S105. Based on the target navigation path, generate corresponding motion control instructions, and control the target robot to perform navigation tasks according to the motion control instructions.

[0164] In this step, the determined target navigation path is converted into motion control commands (such as speed, direction, and turning angle) that the robot's motion actuators can recognize and execute. The control commands are then sent to the robot's motion actuators to drive the robot to autonomously travel along the target navigation path and complete the navigation task.

[0165] In the embodiments of this application, motion control commands are commands used to control the motion state of the robot, including parameters such as linear velocity, angular velocity, and turning angle, which are generated by the robot's computing and processing unit and adapted to the robot's motion execution mechanism (such as motors, wheels, etc.).

[0166] In this way, the abstract path planning results are transformed into the robot's actual physical movement, realizing the implementation of the navigation path and ensuring that the robot can travel accurately and safely according to the planned path.

[0167] Here, the robot can precisely execute actions such as steering, acceleration, deceleration, and straight-line movement according to motion control commands, strictly follow the target navigation path, complete autonomous navigation tasks, and adjust its motion state in a timely manner through replanning based on changes in the local environment, ensuring the continuity and safety of navigation.

[0168] In summary, the method described in this application effectively solves the technical defects of existing robot navigation path planning methods, such as limited map information, unreasonable update mechanism, imperfect cost quantification, and poor coordination between global and local paths. Compared with the prior art, it has the following significant beneficial effects.

[0169] Firstly, a multi-dimensional two-dimensional raster layer (geometry, semantics, terrain) is constructed to achieve a comprehensive expression of environmental information, breaking the limitations of traditional single geometric maps. This enables path planning to avoid obstacles while also taking into account semantic risks and terrain adaptability, significantly improving the safety of navigation in complex environments.

[0170] Secondly, a multi-level global and local map system is designed, combined with a differentiated update mechanism (global incremental update with historical memory and local real-time full update), which ensures the stability of the global map while enabling real-time response to dynamic changes in the local environment, thus balancing computational overhead and map real-time performance.

[0171] Third, establish a multi-dimensional cost quantification system, design targeted cost modeling methods for the three types of layers, and generate a drivability cost map through weighted fusion to provide a scientific and unified decision-making basis for path planning, thereby improving the rationality and efficiency of path planning.

[0172] Fourth, it achieves synergy between global path planning and local optimization. Combined with the replanning trigger mechanism, it not only ensures the stability and continuity of global navigation, but also accurately adapts to the dynamic changes of the local environment, avoids navigation failure, and significantly improves the robot's navigation robustness and adaptability in complex indoor and outdoor dynamic environments.

[0173] The robot navigation path planning method based on multi-level maps provided in this application acquires multi-source perception datasets and extracts multi-source perception information to construct a multi-level, multi-dimensional two-dimensional grid map including geometric obstacle layers, semantic information layers, and terrain information layers. It then generates a drivability cost map, plans a global target path, and performs local optimization. This enables the decoupled expression and collaborative use of the robot's multi-dimensional environmental information at different spatial scales, thereby balancing global path planning and local motion control requirements. It improves the rationality of path planning, navigation stability, and overall safety of the robot in complex environments, achieving a collaborative adaptation between global path stability and local environmental real-time performance.

[0174] Please see Figure 4 , Figure 4 This is a schematic diagram of a robot navigation path planning device based on a multi-level map, provided as an embodiment of this application. Figure 4 As shown, the robot navigation path planning device 400 includes: The data perception module 410 is used to acquire the multi-source perception dataset corresponding to the environment in which the target robot is located, and to extract multi-source perception information from the multi-source perception dataset. The map building module 420 is used to map the multi-source sensing information to a two-dimensional grid space under a unified coordinate system to generate multiple two-dimensional grid layers, and to construct a multi-level global two-dimensional grid map and a multi-level local two-dimensional grid map based on the two-dimensional grid layers; wherein, the two-dimensional grid layers include a geometric obstacle layer, a semantic information layer and a terrain information layer; The map fusion module 430 is used to determine the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information, and to generate a global traversability cost map and a local traversability cost map based on the cost quantization data, the multi-level global two-dimensional raster map and the multi-level local two-dimensional raster map. The navigation decision module 440 is used to plan a global target path in the global drivability cost map using the target cost search method, and to perform local optimization on the global target path based on the local drivability cost map and a preset local cost change threshold to determine the target navigation path. The execution control module 450 is used to generate corresponding motion control commands based on the target navigation path, and control the target robot to perform navigation tasks according to the motion control commands.

[0175] Furthermore, when the data sensing module 410 acquires the multi-source sensing dataset corresponding to the environment in which the target robot is located, and extracts multi-source sensing information from the multi-source sensing dataset, the data sensing module 410 is used to: Obtain a multi-source perception dataset corresponding to the environment in which the target robot is located; wherein, the multi-source perception dataset includes at least visual perception data, semantic recognition data, and terrain perception data; For each type of data in the multi-source sensing dataset, determine whether the time interval step size corresponding to the timestamp of the data type and the preset fusion reference time is less than or equal to the preset time synchronization window. If the time interval step size is less than or equal to the time synchronization window, then this type of data is identified as spatial alignment sensing data. If the time interval step size is larger than the time synchronization window, then according to the pose change relationship of the target robot within the time interval step size, spatiotemporal compensation is performed on this type of data to obtain spatial alignment perception data after compensation for this type of data. Multi-source sensing information is extracted from the spatial alignment sensing data; wherein the multi-source sensing information includes at least geometric obstacle information, semantic category information, and terrain structure information.

[0176] Furthermore, when the map building module 420 is used to construct multi-level global two-dimensional raster maps and multi-level local two-dimensional raster maps based on the two-dimensional raster layer, the map building module 420 is used to: Based on the preset global map raster resolution and local map raster resolution, and the preset global update frequency and local update frequency, a global map and a local map are constructed respectively. A global two-dimensional raster layer is determined in the two-dimensional raster layer, and the global two-dimensional raster layer is merged into the global map. Based on a preset historical memory factor, each raster unit in the merged global map is updated using an incremental fusion update mechanism with historical memory to construct a multi-level global two-dimensional raster map. A local two-dimensional raster layer is determined in the two-dimensional raster layer, and the local two-dimensional raster layer is merged into the local map. Then, each raster unit in the merged local map is updated using an instant update mechanism to construct a multi-level local two-dimensional raster map.

[0177] Furthermore, when the map fusion module 430 determines the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information, the map fusion module 430 is used to: For the geometric obstacle layer, based on the geometric obstacle information in the multi-source sensing information, the occupancy probability corresponding to each grid cell in the geometric obstacle layer is determined; Based on the occupancy probability and the preset safety threshold, the geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer is determined using a preset monotonically increasing function. The geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer is determined as the cost quantization data corresponding to the geometric obstacle layer.

[0178] Furthermore, when determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information, the map fusion module 430 is also used to: For the semantic information layer, determine the category weight corresponding to each semantic category in the semantic category information of the multi-source perception information; Based on the semantic category information, the semantic category to which each grid cell in the semantic information layer belongs is determined, and the category weight corresponding to the semantic category is determined as the basic semantic cost quantization data corresponding to the grid cell; Based on the time modulation factor corresponding to the grid cell and the basic semantic cost quantization data corresponding to the grid cell, the semantic cost quantization data corresponding to the grid cell is determined; The semantic cost quantization data corresponding to each grid cell in the semantic information layer is determined as the cost quantization data corresponding to the semantic information layer.

[0179] Furthermore, when determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information, the map fusion module 430 is also used to: For the terrain information layer, based on the terrain structure information in the multi-source sensing information, the slope value and surface roughness value corresponding to each grid cell in the terrain information layer are determined respectively; Based on the slope value and the surface roughness value, determine the terrain cost quantization data corresponding to each grid cell in the terrain information layer; The terrain cost quantization data corresponding to each raster cell in the terrain information layer is determined as the cost quantization data corresponding to the terrain information layer.

[0180] Furthermore, when the map fusion module 430 generates a global traversability cost map and a local traversability cost map based on the cost quantization data, the multi-level global two-dimensional raster map, and the multi-level local two-dimensional raster map, the map fusion module 430 is used to: Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level global two-dimensional raster map and the preset cost weight parameters, the global traversability cost quantization data corresponding to each raster unit in the multi-level global two-dimensional raster map is determined; Based on the global traversability cost quantification data and the multi-level global two-dimensional grid map, a global traversability cost map is generated; Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level local two-dimensional raster map and the preset cost weight parameters, the local traversability cost quantization data corresponding to each raster unit in the multi-level local two-dimensional raster map is determined. A local accessibility cost map is generated based on the local accessibility cost quantification data and the multi-level local two-dimensional raster map.

[0181] Furthermore, when the navigation decision module 440 performs local optimization on the global target path based on the local navigability cost map and a preset local cost change threshold to determine the target navigation path, the navigation decision module 440 is used to: Based on the global target path, local optimization is performed in the local drivability cost map using the target cost search method. During the local optimization process, when the change in the local mobility cost quantification data corresponding to the local mobility cost map is greater than a preset local cost change threshold, local replanning is triggered; when the global target path is unreachable or blocked for a long time, global replanning is triggered. In response to the global target path, the local optimization is completed, and the target navigation path is determined.

[0182] The robot navigation path planning device based on a multi-level map provided in this application acquires a multi-source perception dataset and extracts multi-source perception information to construct a multi-level, multi-dimensional two-dimensional grid map including a geometric obstacle layer, a semantic information layer, and a terrain information layer. It then generates a drivability cost map, plans a global target path, and performs local optimization. This enables the decoupled expression and collaborative use of the robot's multi-dimensional environmental information at different spatial scales, thereby balancing the needs of global path planning and local motion control. It improves the rationality of the robot's path planning, navigation stability, and overall safety in complex environments, achieving a collaborative adaptation between global path stability and local environmental real-time performance.

[0183] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 5 As shown, the electronic device 500 includes a processor 510, a memory 520, and a bus 530.

[0184] The memory 520 stores machine-readable instructions executable by the processor 510. When the electronic device 500 is running, the processor 510 and the memory 520 communicate via the bus 530. When the machine-readable instructions are executed by the processor 510, they can perform the operations described above. Figure 1 The steps of the robot navigation path planning method based on multi-level maps in the method embodiment shown are specifically implemented in the method embodiment and will not be repeated here.

[0185] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 The steps of the robot navigation path planning method based on multi-level maps in the method embodiment shown are specifically implemented in the method embodiment and will not be repeated here.

[0186] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0187] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0188] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0189] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0190] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0191] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A robot navigation path planning method based on a multi-level map, characterized in that, The method includes: Obtain the multi-source perception dataset corresponding to the environment in which the target robot is located, and extract multi-source perception information from the multi-source perception dataset; The multi-source sensing information is mapped to a two-dimensional grid space under a unified coordinate system to generate multiple two-dimensional grid layers. Based on the two-dimensional grid layers, multi-level global two-dimensional grid maps and multi-level local two-dimensional grid maps are constructed respectively. The two-dimensional grid layers include a geometric obstacle layer, a semantic information layer, and a terrain information layer. Based on the multi-source sensing information, the cost quantization data corresponding to each of the two-dimensional raster layers is determined, and based on the cost quantization data, the multi-level global two-dimensional raster map, and the multi-level local two-dimensional raster map, a global traversability cost map and a local traversability cost map are generated accordingly. A global target path is planned in the global navigability cost map using the target cost search method, and the global target path is locally optimized based on the local navigability cost map and a preset local cost change threshold to determine the target navigation path. Based on the target navigation path, corresponding motion control commands are generated, and the target robot is controlled to perform navigation tasks according to the motion control commands.

2. The method according to claim 1, characterized in that, The step of acquiring the multi-source perception dataset corresponding to the environment in which the target robot is located, and extracting multi-source perception information from the multi-source perception dataset, includes: Obtain a multi-source perception dataset corresponding to the environment in which the target robot is located; wherein, the multi-source perception dataset includes at least visual perception data, semantic recognition data, and terrain perception data; For each type of data in the multi-source sensing dataset, determine whether the time interval step size corresponding to the timestamp of the data type and the preset fusion reference time is less than or equal to the preset time synchronization window. If the time interval step size is less than or equal to the time synchronization window, then this type of data is identified as spatial alignment sensing data. If the time interval step size is larger than the time synchronization window, then according to the pose change relationship of the target robot within the time interval step size, spatiotemporal compensation is performed on this type of data to obtain spatial alignment perception data after compensation for this type of data. Multi-source sensing information is extracted from the spatial alignment sensing data; wherein the multi-source sensing information includes at least geometric obstacle information, semantic category information, and terrain structure information.

3. The method according to claim 1, characterized in that, The construction of multi-level global two-dimensional raster maps and multi-level local two-dimensional raster maps based on the two-dimensional raster layer includes: Based on the preset global map raster resolution and local map raster resolution, and the preset global update frequency and local update frequency, a global map and a local map are constructed respectively. A global two-dimensional raster layer is determined in the two-dimensional raster layer, and the global two-dimensional raster layer is merged into the global map. Based on a preset historical memory factor, each raster unit in the merged global map is updated using an incremental fusion update mechanism with historical memory to construct a multi-level global two-dimensional raster map. A local two-dimensional raster layer is determined in the two-dimensional raster layer, and the local two-dimensional raster layer is merged into the local map. Then, each raster unit in the merged local map is updated using an instant update mechanism to construct a multi-level local two-dimensional raster map.

4. The method according to claim 1, characterized in that, The step of determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information includes: For the geometric obstacle layer, based on the geometric obstacle information in the multi-source sensing information, the occupancy probability corresponding to each grid cell in the geometric obstacle layer is determined; Based on the occupancy probability and the preset safety threshold, the geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer is determined using a preset monotonically increasing function. The geometric cost quantization data corresponding to each grid cell in the geometric obstacle layer is determined as the cost quantization data corresponding to the geometric obstacle layer.

5. The method according to claim 4, characterized in that, The step of determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information further includes: For the semantic information layer, determine the category weight corresponding to each semantic category in the semantic category information of the multi-source perception information; Based on the semantic category information, the semantic category to which each grid cell in the semantic information layer belongs is determined, and the category weight corresponding to the semantic category is determined as the basic semantic cost quantization data corresponding to the grid cell; Based on the time modulation factor corresponding to the grid cell and the basic semantic cost quantization data corresponding to the grid cell, the semantic cost quantization data corresponding to the grid cell is determined; The semantic cost quantization data corresponding to each grid cell in the semantic information layer is determined as the cost quantization data corresponding to the semantic information layer.

6. The method according to claim 5, characterized in that, The step of determining the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information further includes: For the terrain information layer, based on the terrain structure information in the multi-source sensing information, the slope value and surface roughness value corresponding to each grid cell in the terrain information layer are determined respectively; Based on the slope value and the surface roughness value, determine the terrain cost quantization data corresponding to each grid cell in the terrain information layer; The terrain cost quantization data corresponding to each raster cell in the terrain information layer is determined as the cost quantization data corresponding to the terrain information layer.

7. The method according to claim 1, characterized in that, The process of generating a global traversability cost map and a local traversability cost map based on the cost quantification data, the multi-level global two-dimensional raster map, and the multi-level local two-dimensional raster map includes: Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level global two-dimensional raster map and the preset cost weight parameters, the global traversability cost quantization data corresponding to each raster unit in the multi-level global two-dimensional raster map is determined; Based on the global traversability cost quantification data and the multi-level global two-dimensional grid map, a global traversability cost map is generated; Based on the cost quantization data corresponding to each two-dimensional raster layer under the multi-level local two-dimensional raster map and the preset cost weight parameters, the local traversability cost quantization data corresponding to each raster unit in the multi-level local two-dimensional raster map is determined. A local accessibility cost map is generated based on the local accessibility cost quantification data and the multi-level local two-dimensional raster map.

8. The method according to claim 1, characterized in that, The step of performing local optimization on the global target path based on the local navigability cost map and a preset local cost change threshold to determine the target navigation path includes: Based on the global target path, local optimization is performed in the local drivability cost map using the target cost search method. During the local optimization process, when the change in the local mobility cost quantification data corresponding to the local mobility cost map is greater than a preset local cost change threshold, local replanning is triggered; when the global target path is unreachable or blocked for a long time, global replanning is triggered. In response to the global target path, the local optimization is completed, and the target navigation path is determined.

9. A robot navigation path planning device based on a multi-level map, characterized in that, The robot navigation path planning device includes: The data perception module is used to acquire the multi-source perception dataset corresponding to the environment in which the target robot is located, and to extract multi-source perception information from the multi-source perception dataset. The map building module is used to map the multi-source sensing information to a two-dimensional grid space under a unified coordinate system to generate multiple two-dimensional grid layers, and to construct multi-level global two-dimensional grid maps and multi-level local two-dimensional grid maps based on the two-dimensional grid layers; wherein, the two-dimensional grid layers include a geometric obstacle layer, a semantic information layer and a terrain information layer; The map fusion module is used to determine the cost quantization data corresponding to each of the two-dimensional raster layers based on the multi-source sensing information, and to generate a global traversability cost map and a local traversability cost map based on the cost quantization data, the multi-level global two-dimensional raster map and the multi-level local two-dimensional raster map. The navigation decision module is used to plan a global target path in the global drivability cost map using the target cost search method, and to perform local optimization on the global target path based on the local drivability cost map and a preset local cost change threshold to determine the target navigation path. The execution control module is used to generate corresponding motion control commands based on the target navigation path, and control the target robot to perform navigation tasks according to the motion control commands.

10. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. The machine-readable instructions are executed by the processor to perform the steps of the robot navigation path planning method based on a multi-level map as described in any one of claims 1 to 8.

Images