Robot adaptive path planning method and system under incomplete environmental information

Abstract

This invention discloses a method and system for adaptive path planning of robots under incomplete environmental information, relating to the field of robot path planning technology. The method includes: converting the constructed environmental perception information into a bird's-eye view feature map based on a bird's-eye view perspective; constructing a semantic grid using a fully convolutional network; generating a basic global path based on static environmental information using a heuristic search algorithm; integrating an optimized environmental risk entropy model into the heuristic search framework to replan the basic global path based on risk perception; continuously monitoring changes in local risk entropy of the robot's surrounding environment based on a local sliding time window, and constructing performance evaluation indicators to guide local optimization decisions. This method incorporates environmental risk entropy results into the basic global path replanning, dynamic cost map updating, and local sliding time window optimization processes, improving the safety, adaptability, and executability of robot path planning under incomplete environmental information conditions.

Description

Technical Field

[0001] This invention relates to the field of robot path planning technology, and in particular to a method and system for adaptive path planning of robots under conditions of incomplete environmental information. Background Technology

[0002] With the rapid development of robotics technology, its applications in industrial production, logistics, and disaster relief are becoming increasingly widespread. Path planning, as a core technology for autonomous robot navigation, directly impacts the robot's operational efficiency and safety in complex environments. Traditional path planning methods typically rely on complete environmental information, generating feasible paths through global path search algorithms and local path optimization algorithms. However, in practical applications, robots often face challenges such as incomplete environmental perception information, frequent appearance of dynamic obstacles, and rapid environmental changes. This makes it difficult for traditional path planning methods to meet the real-time and robustness requirements of complex scenarios.

[0003] Under conditions of incomplete environmental information, robots can only acquire local environmental information through limited sensors. This leads to a lack of completeness in global path planning, making it prone to generating infeasible paths due to missing information. Furthermore, the motion state of obstacles in dynamic environments is uncertain. Traditional path planning methods typically assume a static or slowly changing environment, making it difficult to effectively handle rapid movement or sudden changes in obstacles. In addition, multi-objective optimization problems in path planning often involve mutual constraints. Traditional methods usually employ fixed-weight objective functions, lacking the ability to adaptively adjust to changes in the dynamic environment, resulting in planning results that fail to meet the needs of multiple objectives. Summary of the Invention

[0004] Based on the technical problems existing in the background technology, this invention proposes a robot adaptive path planning method and system under incomplete environmental information. The optimized environmental risk entropy result of this method is introduced into a heuristic search framework, dynamic cost map and local sliding time window evaluation mechanism to guide the risk perception replanning of the basic global path and the adjustment of local paths, thereby improving the safety, adaptability and feasibility of robot path planning under incomplete environmental information.

[0005] The robot adaptive path planning method proposed in this invention under conditions of incomplete environmental information includes: After preprocessing multi-source sensor data, multimodal features are extracted, and environmental perception information is generated through hierarchical feature fusion. Based on the bird's-eye view, environmental perception information is converted into bird's-eye view feature maps, and a semantic grid is constructed by combining a fully convolutional network. A heuristic search algorithm is used to generate basic global paths based on static environmental information to achieve global environmental modeling. An environmental risk entropy model is constructed to quantify dynamic obstacle threats. The parameters of the environmental risk entropy model are optimized through reinforcement learning to assess local path risks. The optimized environmental risk entropy model is integrated into a heuristic search framework. A global evaluation function for risk perception is constructed based on the geometric length of the path segment, the average risk entropy of the path segment, and the predicted risk entropy of the target direction. The basic global path is then replanned. Furthermore, the local risk entropy changes of the robot's surrounding environment are continuously monitored based on a local sliding time window. When the local risk entropy changes meet preset conditions, local optimization decisions are guided by the set performance evaluation indicators.

[0006] Furthermore, the extraction of multimodal features after preprocessing the multi-source sensor data specifically includes: Preprocessing: Noise suppression and outlier processing of multi-source sensor data based on Kalman filter and wavelet transform, and synchronization and spatial alignment of multi-source sensor data based on timestamp; A feature extraction model based on Vision Transformer is used to extract features from the preprocessed multimodal features. The correlation between visual data, LiDAR point cloud data and millimeter-wave radar detection data in multi-source sensor data is captured by a multi-head self-attention mechanism and a cross-modal attention module.

[0007] Furthermore, the hierarchical feature fusion adopts a multi-scale feature pyramid structure and a channel attention mechanism to generate unified environmental perception information through progressive upsampling and skip connections.

[0008] Furthermore, the process of converting environmental perception information into a bird's-eye view feature map and constructing a semantic grid using a fully convolutional network specifically involves: By projecting visual data from environmental perception information into the bird's-eye view space through inverse perspective mapping, and combining the spatial mapping results of LiDAR point cloud data and millimeter-wave radar detection data, a multi-scale feature pyramid network is used to extract features at different scales and fuse them to obtain the bird's-eye view feature map. The bird's-eye view feature map is processed by a fully convolutional network, which is based on the DeepLabV3+ architecture. Multi-scale contextual information is extracted through the hollow spatial pyramid pooling module, and the semantic segmentation boundary is optimized by combining conditional random fields to generate a semantic raster.

[0009] Furthermore, the generation of a basic global path based on static environment information using a heuristic search algorithm, wherein the heuristic search algorithm employs a hybrid approach... The algorithm, whose cost function is: ; in, For position The comprehensive value of the place For obstacle cost term, For the target distance cost term, For traversable area constraints, These are the weighting coefficients.

[0010] Furthermore, the formula for calculating the risk entropy of the environmental risk entropy model is as follows: ; in, For grid cells The environmental risk entropy value at the location, This is the index number for the risk characteristics. The total number of risk characteristics. For the grid row index and grid column index after the robot's local environment is divided, For the first Risk characteristics in grid cells The probability distribution value in For the first Weighting coefficients for various risk characteristics.

[0011] Furthermore, the reinforcement learning adopts an Actor-Critic architecture based on policy gradients, uses path safety indicators as reward functions, and adjusts the environmental risk entropy model parameters through a proximal policy optimization algorithm.

[0012] Furthermore, the performance evaluation metrics are: ; in, In order to be in Performance evaluation metrics at any time Represents path length. This represents the average risk entropy value of the path. To measure the smoothness of the path, , and These are dynamically adjusted weighting coefficients.

[0013] Furthermore, the optimized environmental risk entropy model is integrated into a heuristic search framework to construct a global evaluation function that integrates path length and environmental risk, used for risk-aware replanning of the basic global path. ; ; ; in, For the current node The global evaluation function, From the starting point to the current node The actual cost function, To start from the current node Heuristic estimation of the cost function to the target point. From the starting point to the current node The total number of path segments traversed. For path segment index, For the first End node on the path segment For the first The starting node on the segment path, Adjacent nodes and The Euclidean distance between them represents the first... The geometric length of the path segment. For risk weighting coefficients, For the first The average risk entropy value of the path segment. For the current node To the target point The Euclidean straight-line distance, As a risk adjustment factor, The predicted risk entropy for the target direction.

[0014] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the method described above.

[0015] The advantages of the robot adaptive path planning method and system under incomplete environmental information provided by this invention are as follows: This embodiment does not only use environmental risk as an evaluation indicator after path generation, but also uses environmental risk entropy results as planning information that can participate in path search and local adjustment. By converting multi-source perception results into a bird's-eye view semantic grid, an environmental expression that can be directly called upon for path search is formed; by constructing an environmental risk entropy model, obstacle distribution density, obstacle motion state, terrain geometric features, and passable area width are uniformly quantified into local risk indicators; by adaptively optimizing the risk feature weights, and introducing the optimized risk entropy results into the global evaluation function, local candidate path generation, dynamic cost map update, and sliding time window path adjustment process, the path planning results can be updated according to changes in local risk. Therefore, the safety, adaptability, and executability of robot path planning under incomplete environmental information conditions can be improved. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 This is a schematic diagram of the global path generation and replanning process based on BEV semantic grids. Figure 3 This is a flowchart of the local candidate path generation and real-time adjustment process based on risk entropy. Detailed Implementation

[0017] The technical solution of the present invention will now be described in detail through specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0018] like Figures 1 to 3 As shown, the robot adaptive path planning method under incomplete environmental information proposed in this invention includes: S1. After preprocessing the multi-source sensor data, extract multimodal features and generate environmental perception information through hierarchical feature fusion. S2. Based on the bird's-eye view, environmental perception information is converted into bird's-eye view feature map, and a semantic grid is constructed by combining a fully convolutional network. A basic global path based on static environmental information is generated through a heuristic search algorithm to realize global environmental modeling. S3. Construct an environmental risk entropy model to quantify dynamic obstacle threats, and optimize the parameters of the environmental risk entropy model through reinforcement learning to assess local path risks. S4. Integrate the optimized environmental risk entropy model into the heuristic search framework, construct a global evaluation function for risk perception based on the geometric length of the path segment, the average risk entropy of the path segment, and the predicted risk entropy of the target direction, and replan the basic global path; and continuously monitor the local risk entropy changes of the robot's surrounding environment based on a local sliding time window, and guide local optimization decisions based on the performance evaluation index constructed according to the path length, average risk entropy, and smoothness when the local risk entropy changes meet the preset conditions.

[0019] Based on the technical problems existing in the background technology, this embodiment proposes a robot adaptive path planning method and system under incomplete environmental information. By combining multimodal perception fusion technology with dynamic environment modeling, and introducing the environmental risk entropy model into the heuristic search and local path optimization process, the safety, adaptability and executability of path planning results under incomplete environmental information are improved.

[0020] This embodiment focuses on resolving the contradiction between incomplete perception and real-time path planning in complex dynamic environments. The overall solution comprises four core modules: multimodal perception data fusion, global environment modeling, risk entropy assessment, and adaptive path optimization, forming a path planning process from environmental perception, global modeling, risk assessment to path optimization. Specifically, step S1 preprocesses, synchronizes, and fuses features from multi-source sensor data to generate unified environmental perception information; step S2 provides environmental constraints for basic global path generation through global semantic modeling and planning based on bird's-eye view (BEV); step S3 introduces an environmental risk entropy model to quantify potential risks in dynamic environments; and step S4 uses the optimized environmental risk entropy model for path replanning and local optimization decisions, improving the safety and adaptability of path planning in complex environments.

[0021] It should be noted that the comprehensive cost map, dynamic cost map, hierarchical dynamic cost map, global hierarchical dynamic cost map, and local cost map mentioned in the following text are all expressions of the cost map at different processing stages or at different spatial scales. Among them, the comprehensive cost map corresponds to the static cost map used in step S2 when generating the basic global path based on the semantic grid; the global hierarchical dynamic cost map corresponds to the planning cost map formed in step S4 after introducing environmental risk entropy and dynamic update mechanism on the basis of the comprehensive cost map; the local cost map is the robot's neighborhood region extracted or updated from the global hierarchical dynamic cost map, used for local path adjustment.

[0022] In one embodiment, step S1 aims to address the noise interference, temporal asynchrony, and spatial alignment issues of multi-source sensor data. Simultaneously, it generates unified environmental perception information through multimodal feature extraction and fusion, laying a solid foundation for subsequent global environment modeling and path planning. To this end, this solution designs a complete multi-source data processing workflow, including data preprocessing, multimodal feature extraction and fusion, and a hierarchical feature fusion network. S1 specifically includes sub-steps S1.1 to S1.4: S1.1 Data preprocessing based on Kalman filter and wavelet transform; To address the noise suppression problem of multi-source sensor data, a multi-scale decomposition using the db4 wavelet basis function is employed. The db4 wavelet basis function is chosen because of its good orthogonality and compact support, which effectively captures the local features of the signal. By setting a four-level wavelet decomposition, the different frequency components of the original signal are separated, with the high-frequency component mainly containing noise information and the low-frequency component retaining the main features of the signal. The basic expression of the wavelet transform is: .

[0023] in, The wavelet transform coefficients represent In scale Translational displacement Lower and mother wavelet functions The degree of similarity; the larger the absolute value of this coefficient, the stronger the signal component at that time-frequency position. For the complex conjugate of the mother wavelet function, The mother wavelet function is the db4 wavelet basis function, which has good orthogonality and tight support, and is used to capture the local features of the signal. This is a time variable, representing consecutive time points in the signal sampling process; The original input signal represents the data sequence collected by multiple sensors or the feature sequence converted from them. For image data, continuous frame images or image feature sequences can be used as the processing object. For point cloud, detection, or state data, the corresponding time series or feature sequences can be used as the processing object.

[0024] The scaling parameter (scaling factor) controls the stretching or compression of the wavelet: The time wavelet is stretched, corresponding to the low-frequency components (the general trend of the signal). Time wavelets are compressed, corresponding to high-frequency components (noise or details). In this embodiment, by setting four levels of wavelet decomposition, different frequency components of the signal are separated at different scales. The translation parameter (position parameter) controls the sliding position of the wavelet on the time axis, and is used to analyze the local characteristics of the signal at different times.

[0025] wavelet transform For multiscale decomposition: by selecting different The values ​​(corresponding to 4-level decomposition) separate the original signal into high-frequency detail components (mainly containing noise) and low-frequency approximation components (mainly containing effective signal). A soft thresholding method is applied to the high-frequency coefficients (the threshold is adaptively adjusted based on the signal's local variance) to suppress noise. After processing, an inverse wavelet transform is performed to reconstruct the denoised signal.

[0026] In this embodiment, a soft thresholding method is used to process high-frequency coefficients during noise suppression. The selection of the soft threshold is adaptively adjusted based on the local variance characteristics of the signal, which ensures the denoising effect while avoiding excessive smoothing of signal details.

[0027] Considering the nonlinear characteristics of sensor data, a state estimation filtering framework based on extended Kalman filtering is designed. This framework continuously optimizes the state estimate through a prediction-update iterative process. The prediction step predicts the state based on the filtering framework and the state estimate result from the previous time step, while the update step corrects the prediction result based on the actual measured value. After the environmental risk entropy model is constructed, the change in local risk entropy can be used as a reference factor for adaptive covariance adjustment. To improve the robustness of the method in this embodiment, an adaptive covariance matrix adjustment mechanism is introduced, enabling the Kalman filter to dynamically adjust its gain coefficient according to data quality.

[0028] A filtering framework for state estimation based on extended Kalman filtering is used for sensor data preprocessing and state estimation. In one specific implementation, the state... It can be set to: .

[0029] in, , Indicates time Below, the robot's position in a planar coordinate system. Indicates time linear velocity, Indicates time The heading angle.

[0030] Control quantity It can be set to: ; in, Indicates time acceleration, Indicates time angular velocity.

[0031] State prediction can be performed as follows: ; ; ; .

[0032] in, For time step, , For a moment Below, the robot's position in a planar coordinate system. For a moment linear velocity, For a moment The heading angle.

[0033] And the Jacobian matrix is ​​calculated based on this state transition model. The state covariance matrix is ​​then updated using prediction: ; in, To predict the covariance matrix, For a moment The covariance matrix, Let be the process noise covariance matrix.

[0034] Outlier detection employs an adaptive threshold mechanism based on statistical characteristics. This mechanism calculates local statistical features of the data using a sliding window and, combined with confidence interval theory, establishes a dynamic threshold judgment standard. When data deviates from the normal range, the system automatically marks and temporarily stores suspicious data points. Through temporal correlation analysis of multiple frames of data, the credibility of outliers is further confirmed, thereby achieving reliable anomaly detection.

[0035] S1.2, Multi-source data synchronization mechanism; Time synchronization is a crucial step in multi-source sensor data fusion. This embodiment employs a hierarchical synchronization architecture based on the NTP protocol. At the hardware level, a high-precision clock source provides a reference time, and each sensor module maintains periodic alignment with the master clock through a time synchronization network. The clock skew estimation model considers the effects of transmission delay and random errors. .

[0036] in, The total time deviation represents the total time difference between the two sensor modules (or between the sensor module and the master clock), and this value is used for subsequent time synchronization correction. Indicates clock skew. Due to network transmission delay, This is random error.

[0037] Multiple sets are obtained through multiple bidirectional timestamp exchanges. The measured value can be estimated simultaneously using the least squares estimation method. and The optimal value (and suppress random error) (This affects the time of the sensor), thereby enabling precise time correction.

[0038] To address the asynchronous sampling problem of different sensor modules, a data alignment mechanism based on a sliding time window is designed. This mechanism establishes a unified time base and uses a dynamically adjusted time window to collect sampling data from each sensor module at similar times. For cases where sampling frequencies are inconsistent, an adaptive interpolation algorithm is used for resampling to ensure that all multi-source sensor data are aligned on a unified time axis. The interpolation algorithm dynamically selects an appropriate interpolation method based on data characteristics, ensuring both alignment accuracy and computational efficiency.

[0039] S1.3 Feature extraction based on Vision Transformer; A Vision Transformer-based feature extraction model is employed, and deep feature learning from multi-source sensor data is achieved through a multi-head self-attention mechanism. Input data from different modalities are preprocessed and then uniformly transformed into a feature space of the same dimension.

[0040] Multi-source sensor data includes visual data, lidar point cloud data, and millimeter-wave radar detection data; visual data is acquired through cameras, lidar point cloud data is collected through lidar sensors, and millimeter-wave radar detection data is used to characterize the target's distance, orientation, or velocity information.

[0041] Visual data is divided into fixed-size image patches. LiDAR point cloud data is converted into a regular grid structure through voxelization. Millimeter-wave radar detection data is spatially mapped based on target distance, azimuth, or velocity information to generate detection feature representations corresponding to the bird's-eye view space. Subsequently, the data for each modality is unified in feature space through an independent linear mapping layer.

[0042] The core of the multi-head self-attention mechanism lies in the parallel computation of multiple sets of attention, each set independently capturing different types of associations between features. Specifically, for the input feature matrix X, it corresponds to the unified feature representation obtained after preprocessing, spatiotemporal alignment, and linear mapping of visual data, LiDAR point cloud data, and millimeter-wave radar detection data (i.e., the features obtained after S1.1 and S1.2). First, it uses three sets of learnable weight matrices... , , Generate a matrix of query, key, and value elements: .

[0043] in, Indicates the first One point of attention, The first Each attention point corresponds to three sets of learnable weight matrices. , , . The first The query matrix, key matrix, and value matrix corresponding to each attention head.

[0044] Attention weights are obtained by calculating the similarity between the query and the key, and then scaling and softmax normalizing. : .

[0045] Wherein, scaling factor This is used to prevent gradient vanishing when the feature dimension is high. The product of the attention weights and the value yields the first... Weighted features of each attention head output .

[0046] The outputs of all attention heads are concatenated and subjected to a linear transformation to obtain the final output of the multi-head attention. : .

[0047] in, For splicing operations, For the first The output of each attention head, For the number of attention heads, The output weight matrix (linear projection matrix) is used to linearly transform the concatenated multidimensional features back to the uniform dimension required by the model, which facilitates subsequent residual connections and layer normalization.

[0048] S1.4, Hierarchical Feature Fusion Network; The final output of multi-head attention in step S1.3 As input to the hierarchical feature fusion network, the network employs a multi-scale pyramid structure to achieve progressive fusion of features at different levels. This network contains five scale layers. Each layer progressively expands its receptive field through downsampling operations (e.g., downsampling factors of 1, 2, 4, 8, and 16) to capture environmental information at different scales. During feature fusion, this is achieved across the five scale layers. In the feature fusion process, channel attention mechanisms are introduced to adaptively adjust the weight allocation of features from different modalities. For feature maps... (Features from a single layer in a multi-scale pyramid), where C is the number of channels (e.g., the number of channels for 5 scales can be set to...). , , , and H×W represents the spatial resolution. and These are feature maps The number of rows and columns is used to calculate channel attention weights through global average pooling and a two-layer fully connected network.

[0049] .

[0050] in, These are channel attention weights, with each channel corresponding to a weight value. The Sigmoid activation function outputs attention weights between 0 and 1. , For the dimensionality-reduced weight matrix (dimension reduction first reduces computational cost). This is an upgraded weight matrix (restored to the original number of channels). It is the ReLU activation function. Feature maps row and column indexes, For feature map at location The eigenvector at that location.

[0051] Attention weight of this channel The process described is as follows: Input feature map Perform global average pooling on each channel. The output vector is then fed into a fully connected network for dimensionality reduction. ) and dimensional ascension ( Processed and output through an activation function. This mechanism highlights important features and suppresses redundant information by learning the dependencies between channels.

[0052] The feature decoder employs a progressive upsampling strategy, preserving fine-grained feature information through skip connections. At each decoding layer, upsampled high-level features are fused with corresponding low-level features to ensure the final feature representation contains both rich semantic information and retains detailed spatial details. To improve the feature fusion effect, an attention-guided feature selection mechanism is introduced during the decoding process. For feature fusion at layer l, the adaptive fusion computation based on the attention map is as follows: .

[0053] in, For the first The feature map output after layer fusion combines high-level and low-level features, containing rich semantic information while preserving spatial details, and can be used by subsequent decoding layers or task heads. For the first The attention map of the layer has the same spatial size as the feature map, with each element taking a value between [0,1], representing the fusion weight of the high-level features at the corresponding spatial location. Generated by network learning, it can adaptively determine the fusion ratio based on the image content. This is element-wise multiplication (Hadamard product). The attention map is then used. Multiply each channel at the corresponding position in the feature map to achieve weight modulation at spatial position. For the first High-level features of a layer typically come from the upsampled output of the layer above the feature decoder or deep features of the encoder. They have a large receptive field and rich semantic information (such as object category and overall structure), but low spatial resolution and significant loss of detail. For the first The low-level features of a layer, which are usually derived from the shallow features of the corresponding layer of the feature encoder (passed through skip connections), have high spatial resolution and rich detail information (such as edges, textures, and corners), but weak semantic information.

[0054] The adaptive fusion calculation formula for attention maps allows this dynamic fusion mechanism to adaptively adjust fusion weights based on the feature importance of different regions, improving the expressive power of the fused features. Ultimately, through progressive fusion of multi-layer features, the system can generate a unified feature representation containing multi-scale environmental information, providing a reliable perceptual foundation for subsequent path planning tasks.

[0055] In one embodiment, the core objective of step S2 is to provide global navigation guidance for the robot through global environment modeling and path planning based on BEV (bird's-eye view). Addressing the issue of missing global semantic information in complex environments, this step generates a bird's-eye view feature map using viewpoint transformation technology and extracts environmental semantic features using a deep learning model to form a global environment modeling result. Based on this, a hybrid heuristic path planning algorithm is designed, comprehensively considering constraints such as static obstacles, traversable areas, and target locations to generate a basic global path that satisfies static environment constraints. S2 specifically includes sub-steps S2.1 to S2.3.

[0056] S2.1, Viewpoint Transformation and Feature Map Generation Based on IPM; Environmental perception information is projected onto the bird's-eye view space through inverse perspective mapping, and features at different scales are extracted and fused using a multi-scale feature pyramid network to obtain the bird's-eye view feature map.

[0057] This step first uses Inverse Perspective Mapping (IPM) technology to convert the forward-looking camera image into a bird's-eye view. First, it establishes the projection relationship between spatial points in the world coordinate system and the image plane, and then uses the inverse transformation of this projection relationship to achieve inverse perspective mapping from the image plane to the ground plane: .

[0058] in, These are the column and row coordinates on the image plane (unit: pixels). For the camera intrinsic parameter matrix, Let be a rotation matrix. It is a translation vector. Let be the camera extrinsic parameter matrix, describing the camera's pose in the world coordinate system. These are three-dimensional spatial coordinates in the world coordinate system.

[0059] Among them, the camera extrinsic matrix This describes the pose relationship of the camera in the world coordinate system: .

[0060] in, This is the camera's rotation angle (pitch angle) around the Y-axis. The height of the camera's optical center above the ground (unit: meter or millimeter).

[0061] In practical implementation, the camera intrinsic parameter matrix was obtained using Zhang's calibration method. A 9×6 checkerboard calibration board was used, and 20 sets of calibration images from different angles were acquired to ensure calibration accuracy. After obtaining the camera parameters, a ground plane assumption was established, and the height was determined. and rotation angle The corresponding coordinate transformation relationship is constructed. Image resampling is implemented based on the OpenCV framework to generate a 512×512 resolution bird's-eye view space covering a 50m×50m area in front. To improve image quality, an adaptive histogram equalization method is used to enhance image contrast, ensuring clear environmental representation under different lighting conditions.

[0062] For LiDAR point cloud data, the point cloud can be projected onto the robot coordinate system or the ground coordinate system based on the sensor extrinsic parameter calibration relationship, and then rasterized according to the bird's-eye view spatial resolution. For millimeter-wave radar detection data, the detection points can be mapped to the same bird's-eye view space based on the target distance, azimuth angle, and radial velocity, and the velocity information can be used as a supplementary feature for dynamic target risk assessment. The bird's-eye view representations from the above different sources are aligned and fused under a unified spatial resolution to obtain a bird's-eye view feature map.

[0063] After obtaining the bird's-eye view space, the environmental representation capability is enhanced through multi-scale feature extraction. Specifically, a three-layer image pyramid is constructed for multi-scale decomposition, and HOG feature descriptors are extracted at each scale to capture gradient and edge information in the environment. Features from different scales are integrated through a weighted fusion strategy, with weight coefficients determined by cross-validation to balance the contributions of features at different scales. Simultaneously, a spatial attention mechanism is introduced to adaptively enhance the feature representation of key environmental regions, improving the discriminative power of the feature map.

[0064] Finally, the feature map is normalized by batch normalization to eliminate feature distribution offset and by mapping the feature values ​​to a unified interval through min-max normalization, outputting a 512×512×64 dimensional standardized feature map as the bird's-eye view feature map.

[0065] This embodiment's feature extraction method based on IPM (Inverse Perspective Mapping) preserves the geometric and textural information of the environment while enhancing the expressive power of features through multi-scale feature extraction and attention enhancement mechanisms, laying the foundation for subsequent semantic segmentation tasks. Through rigorous parameter calibration and feature enhancement strategies, this method can adapt to different scene conditions, ensuring the stability and reliability of the feature map extraction process.

[0066] S2.2 Semantic segmentation and feature extraction based on fully convolutional networks; The bird's-eye view feature map is processed by a fully convolutional network, which is based on the DeepLabV3+ architecture. Multi-scale contextual information is extracted through the hollow spatial pyramid pooling module, and the semantic segmentation boundary is optimized by combining conditional random fields to generate a semantic raster.

[0067] This step employs an improved DeepLabV3+ network architecture to achieve semantic segmentation of the bird's-eye view feature map. The network backbone uses a ResNet-101 structure, and dilated convolutions are introduced to increase the receptive field. The mathematical expression of dilated convolution is: .

[0068] in, This indicates the spatial sampling location of the output feature map. The response value, This is the feature map location index in the dilated convolution operation, used to indicate the sampling position on the output feature map. Indicates the input feature map, For convolution kernel index, For convolution kernel weights, This represents the void ratio.

[0069] void ratio By inserting elements between adjacent elements of the convolution kernel The addition of zeros enables an exponential expansion of the receptive field, allowing the network to acquire a wider range of contextual information without increasing the number of parameters. A key element in the network design is the Atrous Spatial Pyramid Pooling (ASPP) module, which enhances the fully convolutional network's ability to recognize targets at different scales through parallel multi-scale feature extraction and global context encoding. In the decoder design, an attention-based multi-scale feature fusion strategy is introduced. Through an adaptive weight learning mechanism, it achieves dynamic fusion of high- and low-level features, improving the ability to characterize boundary details.

[0070] To improve the generalization performance of the fully convolutional network, diverse data augmentation strategies are employed, including random rotation, scaling, and brightness adjustment. The training process utilizes a cross-entropy loss function with online hard sample mining, dynamically adjusting sample weights to optimize the recognition of difficult regions. Simultaneously, an auxiliary loss branch is introduced to add supervision signals to the intermediate layer feature maps, accelerating network convergence and enhancing feature representation capabilities. Through this design, the fully convolutional network can effectively identify key environmental elements such as passable areas, obstacles, and restricted areas, providing a reliable environmental understanding foundation for subsequent path planning.

[0071] Step S2.3: Path planning based on a hybrid heuristic search algorithm generates a basic global path based on static environment information, thereby realizing global environment modeling; Based on a fully convolutional network, a hybrid heuristic search algorithm is set up to generate a basic global path based on static environment information. First, a comprehensive cost map is constructed, and the cost function is defined as: .

[0072] in, For position The comprehensive value of the place The obstacle cost term is represented by a Gaussian function that maps obstacle distances to cost values; the closer the path is to an obstacle, the higher the cost, ensuring the path maintains a safe distance from obstacles. The distance cost term is defined by Euclidean distance, which measures the distance from the current position to the target point, guiding the path to converge towards the target point. For the passable area constraint, based on the semantic raster results, a smaller value is taken in passable areas, and a larger value is taken in impassable or restricted areas, guiding the path to prioritize safe and passable areas. These are the weighting coefficients, and adjusting the weights can balance the safety, directness, and standardization of the path.

[0073] In one specific implementation, each cost term can be normalized and then take a value in the range [0,1], where the obstacle cost term... and traversable area constraints A higher value indicates a higher level of risk or constraint, and a greater distance to the target cost. Normalization can be performed based on the ratio of the distance from the current node to the target point to the maximum search distance; each weight coefficient ( , and The sum of weights is a non-negative number that can be set according to task requirements and can be normalized to a weighted sum. ) equals 1.

[0074] Based on the cost map, a hybrid approach was designed. The algorithm performs pathfinding. Hybrid The algorithm employs a priority queue data structure and balances path exploration and utilization by dynamically adjusting the heuristic function weights. During node expansion, an adaptive sampling strategy is used to increase sampling density in key areas, improving planning accuracy. To enhance algorithm efficiency, a quadtree-based spatial index structure is introduced to accelerate collision detection and neighborhood search. Simultaneously, an incremental update mechanism enables dynamic maintenance of the cost map, improving the algorithm's adaptability to dynamic environments.

[0075] The final planned path is smoothed using B-spline curves to improve its continuity and executability. The smoothing process considers robot kinematic constraints, adjusting control point positions through iterative optimization to generate a smooth trajectory that satisfies curvature constraints. This heuristic search algorithm, while ensuring planning efficiency, provides a reasonable global reference for subsequent local path optimization through multi-level constraint considerations and optimization strategies.

[0076] In one embodiment, step S3 focuses on risk assessment and local path planning in dynamic environments, aiming to address the shortcomings of traditional methods in dynamic risk assessment. By introducing an environmental risk entropy model based on Shannon entropy, the uncertainty in dynamic environments is quantified, and reinforcement learning is combined to optimize model parameters, thereby improving the safety and adaptability of path planning. S3 specifically includes sub-steps S3.1 to S3.3.

[0077] S3.1 Construct an environmental risk entropy model; To address the issue of low accuracy in assessing potential risks in complex environments, this step constructs an environmental risk entropy model based on Shannon entropy theory. The main reason for choosing Shannon entropy as the assessment basis is its ability to effectively quantify environmental uncertainty and its favorable mathematical properties and physical meaning. The specific implementation process is as follows: Establish a multidimensional feature probability distribution model for diverse environmental characteristics. Divide the local environment into... Each grid cell contains four key features: obstacle distribution density, obstacle motion state, terrain geometry, and passable area width. This partitioning method ensures the integrity of the environmental representation while facilitating subsequent risk quantification calculations. The obstacle distribution density, obstacle motion state, terrain geometry, and passable area width are jointly determined by the environmental perception information generated in step S1 and the semantic raster generated in step S2. The obstacle motion state can be obtained with the assistance of velocity information from multi-frame perception results or millimeter-wave radar detection data.

[0078] For grid cells Its state vector Defined as: .

[0079] in, This indicates the density of obstacle distribution, reflecting the degree of congestion in a local area; This represents the average speed of obstacles and is used to assess dynamic risk. It indicates the elevation of the terrain, reflecting the geometric features of the environment; The four features represent the passable width and reflect motion constraints. Their selection is based on extensive experimental validation and effectively characterizes the main environmental risk factors. To address the issue of inconsistent dimensions among different features, this method introduces feature normalization. By mapping each feature value to a unified [0,1] interval, the comparability of different features in risk assessment is ensured. This approach not only improves the numerical stability of the environmental risk entropy model but also facilitates subsequent risk entropy calculations.

[0080] Based on feature extraction, this embodiment constructs an environmental risk entropy model based on a multilayer perceptron. The multilayer perceptron is chosen as the basic network structure for the environmental risk entropy model due to its powerful nonlinear mapping capability and good generalization performance. The environmental risk entropy model employs a three-layer multilayer perceptron, introducing nonlinear characteristics through the ReLU activation function to ensure that the environmental risk entropy model can learn complex risk features. For grid cells... Its risk entropy is calculated using a modified Shannon entropy formula: .

[0081] in, For grid cells The environmental risk entropy value at the location, This is the index number for the risk characteristics. The total number of risk characteristics. For the grid row index and grid column index after the robot's local environment is divided, For the first Risk characteristics in grid cells The probability distribution value in For the first Weighting coefficients for various risk characteristics This reflects the importance of different characteristics in risk assessment. The initial values ​​are set based on expert experience and are subsequently dynamically adjusted through an adaptive optimization mechanism.

[0082] To enhance the adaptability of the environmental risk entropy model to environmental changes, this embodiment introduces an attention mechanism. The core idea of ​​the attention mechanism is to enable the model to automatically identify and focus on important environmental features. Especially in complex and dynamic environments, this mechanism can significantly improve the accuracy of risk assessment. The attention weights are calculated using the standard scaled dot product attention mechanism, which adaptively adjusts the importance of different features, allowing the model to better handle the relationships between features.

[0083] Through the above design, the environmental risk entropy model quantifies environmental risks, providing a reliable risk assessment basis for subsequent path planning. The advantages of the environmental risk entropy model are twofold: firstly, it improves the comprehensiveness of risk assessment by considering multi-dimensional features; secondly, it enhances the adaptability of the environmental risk entropy model to dynamic environmental changes by adjusting the importance of different risk features using an attention mechanism.

[0084] S3.2 Design an adaptive parameter optimization mechanism; To address the issue of model parameter failure caused by complex and ever-changing environments, this step designs an adaptive parameter optimization mechanism based on policy gradients. The core reason for choosing reinforcement learning as the optimization method is its ability to optimize decision-making policies through continuous interaction with the environment, making it particularly suitable for handling scenarios with dynamically changing environments. Reinforcement learning employs an Actor-Critic architecture based on policy gradients, using path safety as the reward function, and adjusting the environmental risk entropy model parameters through a proximal policy optimization algorithm. The specific implementation process is as follows: To comprehensively characterize the system state, the state space is designed as a combination of the environmental feature vector at the current position and historical decision information. This design considers both immediate environmental information and temporal correlation, helping the system make more rational decisions. The reward function is designed in the form of a multi-objective combination: .

[0085] in, In the state Execute The reward value obtained is used to guide strategy optimization; The environmental state includes the environmental feature vector of the current location (such as obstacle density, dynamic characteristics, terrain, etc.) and historical decision information; This action typically involves adjusting the parameter update strategy for the weights of each feature in the environmental risk entropy model. The average risk entropy of the path (or the average risk entropy of a local area) reflects the overall safety; the smaller the average risk entropy value, the lower the risk. This represents the distance between the robot and the nearest obstacle; a larger distance indicates a higher obstacle avoidance safety margin. The decision time reflects the real-time nature of the system; the longer the time, the lower the reward.

[0086] These are positive weighting coefficients, which respectively control the impact of risk entropy, obstacle avoidance distance, and real-time performance, and can be normalized to satisfy... In security-first scenarios, settings can be configured. ,For example: .

[0087] As a risk entropy sensitive factor, The larger the value, the more significant the impact of the average risk entropy on the reward value. Distance reward sensitivity factor The larger the value, the more significant the impact of changes in the distance between the robot and the nearest obstacle on the reward value. and The sensitivity of the reward function to changes in risk and distance can be set, and a typical value is: .

[0088] The reward function comprises three key components: the average risk entropy of the path. The distance between the robot and obstacles reflects the overall risk level of the path. Used to characterize safety margin and decision time This is used to constrain the system's real-time performance. By combining the above three types of indicators, the reward function can comprehensively balance path safety, obstacle avoidance effectiveness, and real-time performance. Using an exponential form enhances the reward function's sensitivity to changes in risk entropy and obstacle avoidance distance, which is beneficial for guiding the updating of environmental risk entropy model parameters. The weighting coefficients reflect the principle of prioritizing safety, with the weights related to risk entropy... This value can be set to a relatively large value. To improve learning efficiency and stability, this method adopts an Actor-Critic architecture, which combines policy optimization and value estimation. The Actor network is responsible for generating parameter adjustment policies, while the Critic network evaluates the value of the current state. The two work together to complete parameter optimization.

[0089] For the optimization algorithm selection, this method uses Proximal Policy Optimization (PPO). Compared with traditional policy gradient methods, PPO improves training stability by introducing a probability ratio pruning mechanism to limit the policy update magnitude. This mechanism helps avoid drastic fluctuations in the environmental risk entropy model parameters during iterative optimization, thereby improving the stability and executability of the path planning policy.

[0090] To improve the system's learning efficiency, this method innovatively designs a priority-based experience replay mechanism. The core idea of ​​this mechanism is to learn more from experiences of significant value, rather than simply random sampling. Specifically, sample priorities are determined based on the TD error (the deviation between the TD target and the current estimate), with higher-priority samples having a greater probability of being selected. This mechanism significantly improves training efficiency and accelerates model convergence.

[0091] S3.3, Local candidate path generation based on risk entropy constraints; This method generates a set of local candidate paths that meet risk constraints based on the local risk changes output by the environmental risk entropy model. These local candidate paths do not replace the risk-aware replanning results in step S4, but rather serve as candidate inputs for subsequent path replanning and local path adjustment. Therefore, this method can employ an improved Rapidly-exploring Random Tree Star (RAS). This method, as a local candidate path generation method, handles high-dimensional path search problems while maintaining high computational efficiency. To achieve multi-objective optimization, this method designs a comprehensive path cost function: .

[0092] in, Candidate paths The overall cost is such that a smaller value indicates a better path. The path length is usually the sum of the Euclidean distances between adjacent nodes on the path, reflecting the motion efficiency. To measure the smoothness of a path, curvature integral or change in steering angle is usually used to measure it, reflecting the smoothness of the path; the smaller the smoothness, the better. The risk accumulated along the path is obtained by integrating or summing the environmental risk entropy model along the path, reflecting the overall safety of the path; the higher the risk, the greater the cost. These are weighting coefficients that balance the importance of path length, smoothness, and risk value, and can be dynamically adjusted according to the task.

[0093] The cost function will , and By unifying these objectives and appropriately setting weighting coefficients, a balance among multiple goals was achieved. The measurement uses curvature integrals, which not only consider local steering changes but also reflect the overall smoothness, thus helping to generate paths that better match the robot's motion characteristics.

[0094] This method is based on traditional The framework introduces a risk entropy-based constraint mechanism to limit the probability of candidate paths passing through high-risk areas, and improves the system's ability to cope with environmental uncertainties through safety margin design.

[0095] Regarding the sampling strategy, this method develops an adaptive probability sampling mechanism, breaking through the traditional... Limitations of uniform sampling in the algorithm. The sampling probability exhibits a non-linear relationship with the regional risk entropy: .

[0096] in, Indicates improvement The candidate sampling location in the sampling strategy is used to calculate the probability that the candidate location will be selected as a sampling point. Candidate sampling locations z The probability of being selected as a sampling point ranges from 0 to 1, representing the probability of improvement. The probability of this position becoming a new sampling node in the algorithm; The gradient parameter (positive coefficient) controls the steepness of the probability-risk entropy curve. The larger the value, the higher the probability. The faster the jump in the vicinity; The smaller the value, the more gradual the probability change; For position The risk entropy value at a location is calculated by the environmental risk entropy model, reflecting the degree of environmental uncertainty and potential danger at that location; This is the risk entropy threshold, a critical value used to distinguish between high and low risk areas: for example, when hour, (Biased sampling) means that candidate locations are in low-risk areas, and therefore have a higher sampling probability; when hour, (Suppressed sampling). This means that candidate locations are in high-risk areas, resulting in a lower sampling probability; when... hour, That is, the sampling probability is in the middle state.

[0097] This design uses parameters Adjusting the gradient of probability change using a risk entropy threshold. The algorithm dynamically divides high- and low-risk areas, ensuring sampling coverage while improving exploration efficiency in low-risk areas. Step S3 uses a risk entropy model to quantitatively represent environmental risk and combines it with the parameter adaptive optimization mechanism of reinforcement learning to generate a set of local candidate paths that meet risk constraints, providing candidate path input for subsequent risk perception replanning.

[0098] In one embodiment, step S4 further enhances the dynamic adaptability of path planning by combining a risk perception model and a real-time optimization strategy. This step, based on global path planning, monitors environmental changes in real time and dynamically adjusts the path planning strategy, improving the robot's safety and adaptability in navigation tasks in complex dynamic environments. Specifically, step S4 includes sub-steps S4.1 to S4.3.

[0099] S4.1 Construction of a hierarchical dynamic cost map system; Based on the environmental risk entropy model built in S3.1, this step further constructs a hierarchical dynamic cost map system. This system extends the environmental risk entropy model in S3.1 to a multi-scale level, enabling simultaneous support for global path assessment and local path adjustment.

[0100] At the global level, the risk entropy assessment framework in S3.1 is inherited and extended. The basic cost matrix is ​​constructed as follows: .

[0101] in, This is a globally hierarchical dynamic cost map, which is a two-dimensional matrix storing the comprehensive passage cost of each grid cell using a global environment grid as the carrier; the global environment grid is obtained by discretizing the robot's operating space according to a preset spatial resolution. A two-dimensional grid map, where the row and column positions correspond one-to-one with local spatial areas in the actual environment; These represent the number of rows and columns of the two-dimensional grid map, respectively. For grid cells The comprehensive cost value is used in path planning algorithms for node expansion and path cost evaluation; These are the row and column indices for the grid cells.

[0102] C is not a standalone sensor observation result, but a planning cost map formed by further integrating obstacle distance cost, traffic area constraints and dynamic update mechanism on the basis of the S3.1 environmental risk entropy calculation result.

[0103] Unlike S3.1, which directly uses risk entropy, this step designs a comprehensive cost evaluation equation: .

[0104] in, For grid cells The modified risk entropy value, taking into account the characteristics of risk space propagation, reflects the uncertainty and degree of danger of this unit; For grid cells The obstacle distance cost is usually defined as the Euclidean distance to the nearest obstacle (after normalization or Gaussian transformation), and the closer the distance, the greater the cost. This is a weighting coefficient used to balance the importance of the risk entropy term and the obstacle distance term. It can be dynamically adjusted according to the scenario (e.g., increasing the size of high-risk areas). ).

[0105] The comprehensive cost evaluation equation overcomes the local optimum problem that may be caused by relying solely on risk entropy in S3.1.

[0106] Specifically, spatial correlation was considered in addition to the original risk entropy calculation. The risk entropy calculation formula in S3.1 was improved as follows: .

[0107] in, To determine the total number of risk sources affecting the current grid cell (such as the number of surrounding obstacles or dynamic targets). For risk source index (the first (an obstacle or dynamic target) For the first The probability distribution value of a risk source (usually obtained based on sensor confidence or motion prediction) represents the likelihood of the existence of that risk source. Let be the distance decay function, which is about the th Distance from each risk source to the current grid cell The monotonically decreasing function has a smaller weight as the distance increases, reflecting the spatial propagation characteristics of risk. For the first From risk sources to grid units The Euclidean distance (or path distance).

[0108] The improved risk entropy calculation formula includes a newly added distance decay function. This reflects the spatial propagation characteristics of risk, which is an important supplement to the risk assessment model in S3.1, enabling the system to more accurately describe the spatial distribution patterns of risk.

[0109] Regarding the local dynamic update mechanism, this step innovatively applies the adaptive parameter optimization concept from S3.2 to the cost map update strategy. An adaptive update mechanism based on environmental uncertainty is designed: .

[0110] in, In time and spatial location The adaptive update step size (or cost map update interval) at a given location: the smaller the value, the more frequent the updates; the larger the value, the sparser the updates. Update the step size (constant) based on the baseline. The adaptive adjustment coefficient (positive number) controls the degree to which environmental uncertainty amplifies the step size; To quantify environmental uncertainty, we quantify the degree of dynamic change in the current local environment (such as the average velocity of obstacles, the time gradient of risk entropy, etc.). When the environment is highly uncertain, This reduces the frequency of local cost map updates, allowing for timely reflection of changes in obstacle and risk distribution; when environmental changes are relatively stable, Increase it appropriately to reduce the computational burden.

[0111] Through this multi-layered design, this step expands the scope of the environmental risk entropy model in path planning while inheriting the theoretical foundation of S3. The hierarchical architecture improves the cost map's ability to represent environmental risks at different scales, and the dynamic update mechanism enhances the system's responsiveness to changes in local risks, enabling subsequent path planning processes to simultaneously consider global travel costs and local dynamic risks.

[0112] S4.2 Global Path Dynamic Evaluation and Optimization; This step is the core step in incorporating environmental risk entropy results into path planning decisions. Unlike heuristic searches that expand nodes solely based on path length or obstacle distance, this step introduces the geometric length of the path segment, the average risk entropy of the path segment, and the predicted risk entropy of the target direction into the global evaluation function. This makes the node expansion process constrained by both geometric distance and environmental risk distribution, thereby achieving risk perception and replanning of the basic global path. To this end, an improved global evaluation function is proposed.

[0113] The geometric length of a path segment is calculated from the Euclidean distance between adjacent path nodes in the candidate path or the basic global path.

[0114] The average risk entropy of a path segment is obtained by averaging the risk entropy values ​​of the grid cells covered or traversed by the corresponding path segment using the environmental risk entropy model. .

[0115] in, For the first The average risk entropy value of the path segment. For the first The set of grid cells covered or traversed by the segment path. For grid cells The environmental risk entropy value at the location.

[0116] The target direction prediction risk entropy is determined by the risk entropy prediction value of the candidate extended region or neighboring grid of the current node in the direction of the target node. .

[0117] in, For the current node Predicted risk value for the target direction. Indicates the current node Towards the target node Candidate extended regions or neighborhood grid sets in the direction, The set contains the number of grid cells. For grid cells The environmental risk entropy value at the location is calculated by the environmental risk entropy model in step S3.1.

[0118] Therefore, the geometric length of the path segment, the average risk entropy of the path segment, and the predicted risk entropy in the target direction are determined by the path node coordinates, the grid risk entropy distribution output by the environmental risk entropy model, and the predicted risk entropy in the neighborhood of the current node towards the target direction, respectively. These three factors jointly participate in the actual cost term and heuristic estimation term of the global evaluation function, ensuring that the node expansion process simultaneously reflects both path length and environmental risk. The formula for the global evaluation function is as follows: .

[0119] in, For the current node The global evaluation function, From the starting point to the current node The actual cost function, To start from the current node To the target node The heuristic estimation of the cost function.

[0120] Includes traditional distance metrics and risk assessment elements: .

[0121] in, From the starting point to the current node The total number of path segments traversed. For path segment index, For the first End node on the path segment For the first The starting node on the segment path, Adjacent nodes and The Euclidean distance between them represents the first... The geometric length of the path segment. For risk weighting coefficients, For the first The average risk entropy value of the path segment.

[0122] The impact of risk factors was also taken into account: .

[0123] in, For the current node To the target point The Euclidean straight-line distance, The predicted risk entropy for the target direction. As a risk adjustment factor; In actual implementation, the system first performs a risk entropy assessment, calculating the risk entropy distribution of a local area based on multi-source sensor data and environmental characteristic analysis. Subsequently, a global evaluation function is used. The calculation integrates distance cost and risk factors in a weighted manner to achieve comprehensive optimization of route planning. Risk weight coefficient. The system dynamically adjusts through an adaptive mechanism, automatically updating based on environmental complexity and task requirements to ensure the rationality of planning decisions. Ultimately, the system outputs real-time global risk assessment results. .

[0124] in, This is the global risk entropy value (average risk entropy per unit length) for the entire path. The smaller the value, the lower the overall risk of the path. It is used to evaluate the relative safety of different candidate paths and can also be used as a feedback signal to update the environmental risk entropy model. The total length of the path. For the first The geometric length of a path segment (usually Euclidean distance) reflects the physical proportion of that segment in the path; For the first The average risk entropy value of the path segment. This is calculated by integrating the environmental risk entropy model over the path segment or by applying the risk entropy of the grid cells. The average value is used to characterize the environmental threat level of this route segment. For the first The risk-weighted contribution of a road segment (length × average risk entropy) reflects that even if the risk density is low, a long road segment may still accumulate a large risk.

[0125] The comprehensive assessment results This is used to update the environmental risk distribution map during the path planning process and serves as the evaluation basis for subsequent path replanning. This design, while considering the global path geometric cost, introduces a quantitative assessment and dynamic update mechanism for risk entropy, improving the system's adaptability and safety in complex dynamic environments.

[0126] S4.3 Real-time adjustment and execution of local paths; To address the real-time changing characteristics of risk distribution in the robot's surrounding environment, this step establishes a complete local path dynamic optimization mechanism. By continuously monitoring the degree of risk change in the robot's surrounding environment through a local sliding time window, the average change in local risk entropy is calculated using the following formula: ; .

[0127] in, Deadline Local sliding time window, The number of sampling times contained in this window, and ; For the time index within the window; and These are the local risk entropy values ​​at adjacent time points; For a moment The corresponding average change in local risk entropy; the larger the value, the more drastic the recent changes in local environmental risk.

[0128] By calculating the average change in local risk entropy output by the optimized environmental risk entropy model within a local sliding time window, the change in local risk entropy is obtained. When the local risk entropy value is greater than a preset risk threshold, or the average rate of change of the local risk entropy value is greater than a preset change threshold, local optimization decisions are guided by set performance evaluation indicators, enabling the system to capture the dynamic characteristics of the environment in a timely manner. When the local risk entropy value or average change exceeds the corresponding preset threshold, local path optimization or replanning is triggered.

[0129] In general inspection or mobile robot scenarios, the preset risk threshold can be set to... In scenarios with stricter security constraints, the preset risk threshold can be set to... The preset change threshold can be set to... .

[0130] In another specific implementation, the preset risk threshold can be adaptively determined based on historical operational data. Specifically, the mean of the local average risk entropy can be calculated within a recent time window. and standard deviation The preset risk threshold is determined according to the following formula. : ; in, This is the risk threshold adjustment coefficient, which can be taken as... .

[0131] Similarly, the mean of the average rate of change of local risk entropy can be statistically analyzed within a recent time window. and standard deviation And determine the preset change threshold according to the following formula: ; in, The threshold adjustment coefficient can be taken as... .

[0132] Average rate of change of local risk entropy It can be represented as: ; in, Indicates a sliding time window. This represents the time index within the sliding time window. To carve Centered on (or up to time) The local sliding time window of a time window is a set of discrete time points. They are time points The calculated environmental risk entropy value (output by the risk entropy model in step S3.1 based on the sensor data at that moment).

[0133] Local path optimization or replanning is triggered when any of the following conditions are met: or .

[0134] In the process of local path optimization, in order to simultaneously meet the requirements of local motion efficiency, immediate safety, and motion continuity, a set of local comprehensive performance evaluation indicators was designed to guide local optimization decisions: .

[0135] in, In order to be in Performance evaluation metrics at any time Represents path length. This represents the average risk entropy value of the path. To measure the smoothness of the path, , and These are dynamically adjusted weighting coefficients.

[0136] Using the above strategies, when local risk increases or risk changes more rapidly, the weight of the average risk entropy term is increased; when the risk is low and the path changes smoothly, the weights of the path length and smoothness terms are increased.

[0137] This evaluation metric balances three key factors: It reflects the efficiency of the movement; This reflects the immediate security requirements; Used to characterize the continuity of motion. Weighting coefficients , and By dynamically adjusting through a local adaptive mechanism, the system can make optimal decisions based on the specific characteristics of the local scenario. The unique feature of this local dynamic optimization mechanism is that it achieves a dynamic balance between local safety and efficiency. When the local environmental risk is low, the system prioritizes local path length and smoothness to improve local motion efficiency; conversely, when facing high-risk local situations, the system automatically increases the weight of local safety to ensure the robot's immediate safe operation. Through this local adaptive optimization strategy, the system can not only cope with dynamic changes in the surrounding environment but also maintain high local task efficiency while ensuring local safety.

[0138] This embodiment does not focus on using a single feature extraction network, semantic segmentation network, or path search algorithm, but rather on integrating an environmental risk entropy model throughout the path planning process. Specifically, steps S1 and S2 are used to form an environmental representation that can be used for path search; step S3 is used to convert the obstacle distribution, motion state, terrain geometry features, and passable area width in the local environment into environmental risk entropy, and adaptively optimize the risk feature weights; step S4 further embeds the optimized risk entropy result into the global evaluation function, dynamic cost map, and local sliding time window evaluation mechanism, enabling the path planning process to adjust according to changes in local risk. Through the above steps, this embodiment can improve the safety, adaptability, and executability of path planning results under conditions of incomplete environmental information and dynamic obstacle changes.

[0139] Based on the above description of the embodiments, those skilled in the art will understand that the robot adaptive path planning method and system under incomplete environmental information described in this embodiment can be implemented in pure software or deployed and run on a general-purpose or dedicated computing hardware platform. Based on this essence, the technical solution of this embodiment can be specifically implemented in the form of a software product containing program instructions. This software product can be stored on various non-volatile storage media or directly deployed as a local or cloud service. The program instructions are used to cause computer devices with processing capabilities—including but not limited to personal computers, server clusters, mobile terminals, or other network devices—to execute the steps described in this embodiment.

[0140] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A robot adaptive path planning method under incomplete environmental information, characterized in that, include: After preprocessing multi-source sensor data, multimodal features are extracted, and environmental perception information is generated through hierarchical feature fusion. Based on the bird's-eye view, environmental perception information is converted into bird's-eye view feature maps, and a semantic grid is constructed by combining a fully convolutional network. A heuristic search algorithm is used to generate basic global paths based on static environmental information to achieve global environmental modeling. An environmental risk entropy model is constructed to quantify dynamic obstacle threats. The parameters of the environmental risk entropy model are optimized through reinforcement learning to assess local path risks. The optimized environmental risk entropy model is integrated into a heuristic search framework. A global evaluation function for risk perception is constructed based on the geometric length of the path segment, the average risk entropy of the path segment, and the predicted risk entropy of the target direction. The basic global path is then replanned. Furthermore, the local risk entropy changes of the robot's surrounding environment are continuously monitored based on a local sliding time window. When the local risk entropy changes meet preset conditions, local optimization decisions are guided by the set performance evaluation indicators.

2. The method according to claim 1, characterized in that, The process of preprocessing multi-source sensor data and extracting multimodal features specifically involves: Preprocessing: Noise suppression and outlier processing of multi-source sensor data based on Kalman filter and wavelet transform, and synchronization and spatial alignment of multi-source sensor data based on timestamp; A feature extraction model based on Vision Transformer is used to extract features from the preprocessed multimodal features. The correlation between visual data, LiDAR point cloud data and millimeter-wave radar detection data is captured through a multi-head self-attention mechanism and a cross-modal attention module.

3. The method according to claim 1, characterized in that, The hierarchical feature fusion adopts a multi-scale feature pyramid structure and channel attention mechanism, and generates unified environmental perception information through progressive upsampling and skip connections.

4. The method according to claim 1, characterized in that, The process of converting environmental perception information into a bird's-eye view feature map and constructing a semantic raster using a fully convolutional network is as follows: By projecting visual data from environmental perception information into the bird's-eye view space through inverse perspective mapping, and combining the spatial mapping results of LiDAR point cloud data and millimeter-wave radar detection data, a multi-scale feature pyramid network is used to extract features at different scales and fuse them to obtain the bird's-eye view feature map. The bird's-eye view feature map is processed by a fully convolutional network, which is based on the DeepLabV3+ architecture. Multi-scale contextual information is extracted through the hollow spatial pyramid pooling module, and the semantic segmentation boundary is optimized by combining conditional random fields to generate a semantic raster.

5. The method according to claim 1, characterized in that, The process generates a basic global path based on static environment information using a heuristic search algorithm, wherein the heuristic search algorithm employs a hybrid approach. The algorithm, whose cost function is: ； in, For position The comprehensive value of the place For obstacle cost term, For the target distance cost term, For traversable area constraints, These are the weighting coefficients.

6. The method according to claim 1, characterized in that, The formula for calculating the risk entropy of the environmental risk entropy model is as follows: ； in, For grid cells The environmental risk entropy value at the location, This is the index number for the risk characteristics. The total number of risk characteristics. For the grid row index and grid column index after the robot's local environment is divided, For the first Risk characteristics in grid cells The probability distribution value in For the first Weighting coefficients for various risk characteristics.

7. The method according to claim 1, characterized in that, The reinforcement learning adopts an Actor-Critic architecture based on policy gradients, uses path safety index as reward function, and adjusts the environmental risk entropy model parameters through a near-end policy optimization algorithm.

8. The method according to claim 1, characterized in that, The performance evaluation metrics are: ； in, In order to be in Performance evaluation metrics at any time Represents path length. This represents the average risk entropy value of the path. To measure the smoothness of the path, , and These are dynamically adjusted weighting coefficients.

9. The method according to claim 1, characterized in that, The optimized environmental risk entropy model is integrated into a heuristic search framework to construct a global evaluation function that integrates path length and environmental risk, which is used for risk-aware replanning of the basic global path. ； ； ； in, For the current node The global evaluation function, From the starting point to the current node The actual cost function, To start from the current node Heuristic estimation of the cost function to the target point. From the starting point to the current node The total number of path segments traversed. For path segment index, For the first End node on the segment path, For the first The starting node on the segment path, Adjacent nodes and The Euclidean distance between them represents the first... The geometric length of the path segment. For risk weighting coefficients, For the first The average risk entropy value of the path segment. For the current node To the target point The Euclidean straight-line distance, As a risk adjustment factor, The predicted risk entropy for the target direction.

10. A computer system comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the method according to any one of claims 1-9.

Images