Robot indoor exploration method based on lightweight neural network prediction model

Abstract

The invention relates to a robot indoor exploration method based on a lightweight neural network prediction model, and the method comprises the steps: obtaining a local observation map of a target robot, inputting the local observation map into a preset local prediction model to obtain a local prediction map, and updating a global probability map according to the local prediction map, dividing a free area and an uncertain area according to the updated global probability map, if boundary points of the free area and / or boundary points of the uncertain area exist, clustering all the boundary points, calculating the distance between the target robot and the centroid of each cluster, determining a target point and an exploration path, and after the robot reaches the target point, searching the target robot according to the exploration path. And re-executing the steps of acquiring the local prediction map and determining a new target point until the global map is constructed. Therefore, the problems that autonomous exploration of the robot is poor in performance in a complex environment, the calculation burden is too heavy, and the exploration efficiency is too low are solved, the calculation burden can be reduced, and the autonomous exploration performance is improved.

Description

Technical Field

[0001] This application relates to the field of robotics technology, and in particular to a method for indoor robotic exploration based on a lightweight neural network prediction model. Background Technology

[0002] Autonomous exploration is a key issue in the field of robotics, with important applications ranging from indoor navigation to planetary exploration. It involves creating environmental models through recursive relationships between perception and motion, a process fraught with complex challenges, including the need for real-time decision-making, the variability of the environment, and the construction of reliable environmental models.

[0003] Related technologies rely on boundaries for autonomous exploration, typically employing a greedy strategy. In low-complexity environments, this approach is usually sufficient for most exploration tasks. However, in more information-rich environments, these boundaries fail to effectively guide autonomous exploration. To overcome environmental limitations, related technologies incorporate machine learning techniques to predict map uncertainties and navigate to these uncertain areas, thereby improving the robot's exploration capabilities.

[0004] However, while machine learning-based exploration methods take advantage of the uncertainties in map prediction, they do not consider the errors introduced by the sensor's field of view (FOV). In addition, methods that use large models to predict the overall environment often ignore the computational constraints of the robot's onboard hardware, which may lead to excessive computational burden and thus reduce efficiency. Summary of the Invention

[0005] This application provides a robot indoor exploration method based on a lightweight neural network prediction model to solve the problems of poor performance, excessive computational burden, and low exploration efficiency of robot autonomous exploration in complex environments. This application combines the prediction of unobserved areas with a real-time exploration strategy to reduce the computational burden and improve the performance of autonomous exploration.

[0006] The first aspect of this application provides a robot indoor exploration method based on a lightweight neural network prediction model, comprising the following steps:

[0007] Obtain a local observation map of the target robot;

[0008] The local observation map is input into a preset local prediction model to obtain a local prediction map, and the global probability map is updated according to the local prediction map. The free region and the uncertain region are divided according to the updated global probability map.

[0009] Determine whether there are boundary points of the free region and / or the uncertain region. If there are boundary points of the free region and / or the uncertain region, cluster all boundary points, calculate the distance between the target robot and the centroid of each cluster, determine the target point based on the distance between the target robot and the centroid of each cluster, and determine the exploration path based on the target point. After the robot reaches the target point, re-execute the steps of acquiring the local prediction map and determining the new target point until the global map is constructed.

[0010] Optionally, in some embodiments, the step of dividing the free region and the uncertain region according to the updated global probability map includes:

[0011] Based on the global probability map, regions in the global probability map that are less than a first preset threshold are classified as the free regions;

[0012] The regions in the global probability map that are greater than a second preset threshold are classified as obstacle regions, wherein the second preset threshold is greater than the first preset threshold;

[0013] The regions in the global probability map that are greater than or equal to the first preset threshold and less than or equal to the second preset threshold are classified as the uncertain regions.

[0014] Optionally, in some embodiments, before inputting the local observation map into a preset local prediction model to obtain a local prediction map, the following steps are included:

[0015] Obtain the local map prediction dataset;

[0016] The neural network model is trained using the local map prediction dataset to obtain a trained neural network model, and the trained neural network model is optimized using a preset loss function to obtain the preset local prediction model.

[0017] Optionally, in some embodiments, the preset loss function is:

[0018]

[0019] in, Let m be the predicted map and w be the actual map. mse The mean squared error loss weight value is... For mean squared error loss, w feat The weight values ​​for feature reconstruction loss. The feature reconstruction loss is used.

[0020] Optionally, in some embodiments, updating the global probability map based on the local prediction map includes:

[0021] Based on a preset global probability map update calculation formula, the global probability map is updated according to the local prediction map. The preset global probability map update calculation formula is as follows:

[0022]

[0023] in, Let be the global probability map at time t. Let be the x-coordinate of any location on the global probability map at time t. Let be the ordinate of any location in the global probability map at time t, and α be the update weight value for updating the global probability map from the local prediction map. Let be the local predicted map at time t, where i is the x-coordinate of the predicted map and j is the y-coordinate of the predicted map. This is the global probability map at time t-1.

[0024] A second aspect of this application provides a robotic indoor exploration device based on a lightweight neural network prediction model, comprising:

[0025] The acquisition module is used to acquire a local observation map of the target robot;

[0026] The prediction module is used to input the local observation map into a preset local prediction model to obtain a local prediction map, update the global probability map according to the local prediction map, and divide the free region and uncertain region according to the updated global probability map.

[0027] The exploration module is used to determine whether there are boundary points of the free region and / or the uncertain region. If there are boundary points of the free region and / or the uncertain region, all boundary points are clustered, the distance between the target robot and the centroid of each cluster is calculated, and the target point is determined based on the distance between the target robot and the centroid of each cluster. The exploration path is determined based on the target point. After the target robot reaches the target point, the steps of acquiring the local prediction map and determining the new target point are re-executed until the global map is constructed.

[0028] Optionally, in some embodiments, the prediction module includes:

[0029] The first division unit is used to divide the regions in the global probability map that are less than a first preset threshold into the free regions based on the global probability map.

[0030] The second division unit is used to divide the area in the global probability map that is greater than a second preset threshold into an obstacle area, wherein the second preset threshold is greater than the first preset threshold.

[0031] The third division unit is used to divide the regions in the global probability map that are greater than or equal to the first preset threshold and less than or equal to the second preset threshold into the uncertain regions.

[0032] Optionally, in some embodiments, before inputting the local observation map into a preset local prediction model to obtain a local prediction map, the prediction module includes:

[0033] The acquisition unit is used to acquire the local map prediction dataset;

[0034] The optimization unit is used to train the neural network model using the local map prediction dataset to obtain the trained neural network model, and to optimize the trained neural network model using a preset loss function to obtain the preset local prediction model.

[0035] Optionally, in some embodiments, the preset loss function is:

[0036]

[0037] in, Let m be the predicted map and w be the actual map. mse The mean squared error loss weight value is... For mean squared error loss, w feat The weight values ​​for feature reconstruction loss. The feature reconstruction loss is used.

[0038] Optionally, in some embodiments, updating the global probability map based on the local prediction map includes:

[0039] Based on a preset global probability map update calculation formula, the global probability map is updated according to the local prediction map. The preset global probability map update calculation formula is as follows:

[0040]

[0041] in, Let be the global probability map at time t. Let be the x-coordinate of any location on the global probability map at time t. Let be the ordinate of any location in the global probability map at time t, and α be the update weight value for updating the global probability map from the local prediction map. Let be the local predicted map at time t, where i is the x-coordinate of the predicted map and j is the y-coordinate of the predicted map. This is the global probability map at time t-1.

[0042] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the robot indoor exploration method based on a lightweight neural network prediction model as described in the above embodiments.

[0043] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the robot indoor exploration method based on a lightweight neural network prediction model as described in the above embodiments.

[0044] Therefore, a local observation map of the target robot is acquired, and this local observation map is input into a preset local prediction model to obtain a local prediction map. The global probability map is then updated based on the local prediction map. The updated global probability map is used to divide the region into free and uncertain areas. If boundary points exist for both free and uncertain areas, all boundary points are clustered. The distance from the target robot to the centroid of each cluster is calculated, and a target point is determined based on this distance. An exploration path is then determined based on the target point. After the robot reaches the target point, the steps of acquiring the local prediction map and determining a new target point are repeated until the global map is constructed. This solves the problems of poor performance, excessive computational burden, and low exploration efficiency in autonomous robot exploration in complex environments. This application combines unobserved area prediction with a real-time exploration strategy to reduce computational burden and improve autonomous exploration performance.

[0045] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0046] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0047] Figure 1 This is a schematic diagram of a multi-machine exploration scenario for related technology one;

[0048] Figure 2 A schematic diagram illustrating the generation of small missing unknown spaces in a related technology;

[0049] Figure 3 This is a schematic diagram illustrating the technical principle of related technology two;

[0050] Figure 4 This is a flowchart of a robot indoor exploration method based on a lightweight neural network prediction model provided in an embodiment of this application;

[0051] Figure 5 This is a schematic diagram illustrating the principle of a preset local prediction model provided according to an embodiment of this application;

[0052] Figure 6 This is a schematic diagram illustrating the principle of region division according to an embodiment of this application;

[0053] Figure 7 This is a schematic diagram illustrating the principle of global map construction according to an embodiment of this application;

[0054] Figure 8 This is a schematic diagram showing the results of a comparison experiment of the loss function provided according to an embodiment of this application;

[0055] Figure 9 This is a schematic diagram illustrating the evaluation results based on performance metrics according to one embodiment of this application;

[0056] Figure 10 This is a schematic diagram of a visual comparison result provided according to an embodiment of this application;

[0057] Figure 11 This is a schematic diagram illustrating the evaluation results based on performance metrics according to another embodiment of this application;

[0058] Figure 12 This is a comparative schematic diagram of map reconstruction results provided according to one embodiment of this application;

[0059] Figure 13 This is a schematic diagram illustrating the comparison results of exploration coverage according to an embodiment of this application;

[0060] Figure 14 This is a schematic diagram illustrating the comparison results of exploration efficiency provided according to an embodiment of this application;

[0061] Figure 15 This is a block diagram of a robot indoor exploration device based on a lightweight neural network prediction model according to an embodiment of this application;

[0062] Figure 16 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. Detailed Implementation

[0063] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0064] Before introducing the robot indoor exploration method based on a lightweight neural network prediction model according to the embodiments of this application, let's first introduce the robot exploration methods in related technologies.

[0065] Effective autonomous exploration strategies typically focus on selecting the most informative areas for exploration. This involves estimating or predicting unexplored areas while considering the influence of sensor and environmental factors. Many real-world environments used for exploration exhibit a degree of structure, predictability, and repeatability in their geometric layout. In particular, indoor environments that match these characteristics, such as office buildings or hospitals, often consist of repetitive rooms and corridors. By observing local information in one area of ​​the environment, information about that area and even other areas can be predicted. This inherent predictability improves exploration efficiency, enabling robots to use local observations to guide navigation and mapping.

[0066] like Figure 1 As shown, the related technology proposes and verifies a frontier-based multi-robot autonomous exploration method for efficiently building maps in unknown indoor environments. By continuously moving robots towards the frontier region, the known area is expanded, gradually constructing an occupancy grid for the unknown environment. Each cell in the grid map stores the probability of that area being occupied. Based on prior values, cells are categorized into three states: Open, Occupied, and Unknown. The frontier region is defined as a cluster of adjacent regions between Open Cells and Unknown Cells. Under multi-robot autonomous exploration, each robot maintains an independent local map but broadcasts local perception information. Then, log-odds probabilities are used to additively merge the local maps of multiple robots, updating the overall map in real time. Robot behavior is autonomous, and redundant exploration is avoided by sharing perception data.

[0067] However, as Figure 2 As shown, during the exploration process, some areas may not be fully scanned by the radar, resulting in some unknown spaces. To determine the true situation of these unknown spaces, it may be necessary to repeat the entire multi-robot autonomous exploration process multiple times, which consumes a lot of time and energy. The energy and time required are seriously inconsistent with the expected exploration area.

[0068] like Figure 3As shown, Related Technology 2 proposes a novel framework for robot exploration—MapEx—specifically designed for structured indoor environments. It improves traditional exploration methods by utilizing predicted maps and probabilistic information gain. The system uses a deep learning network for image inpainting (image reconstruction)—LaMa (Large Mask Image Inpainting)—to predict indoor environment maps, particularly for filling in areas unobserved due to sensor limitations (such as the perception range of LiDAR). In the MapEx framework, the input to LaMa is an observed map containing unknown areas, while its output is a complete predicted map. The model infers the content of missing areas based on the image context. To estimate the uncertainty of unknown areas and the reliability of the predictions, MapEx maintains a set of independent LaMa networks, each generating a prediction based on the same observed map. These predictions generate a variance map to measure the prediction uncertainty for different areas. The system then estimates the perceptible area at a given viewpoint based on the mean and variance of the predicted maps, and calculates the information gain to help the robot select the most promising areas for exploration.

[0069] However, while using LaMa deep learning networks for map prediction improves the efficiency of exploration in complex environments, it ignores the computational overhead associated with neural network inference, resulting in a large system size, high computational cost, and poor robustness.

[0070] In summary, the relevant technologies have the following problems: (I) Boundary-based multi-robot autonomous exploration methods identify the boundaries between explored and unexplored areas and guide robots to move towards these boundaries. Multi-robots use collaborative distributed map sharing and autonomous navigation decision-making to explore unknown spaces. However, during the exploration process, obstacle occlusion effects caused by path planning and performance issues of sensor devices result in numerous unexplored unknown areas. If relying on boundary-based autonomous exploration methods, determining the environmental structure of unknown spaces requires repeatedly performing the entire multi-robot autonomous exploration process, which involves re-exploring and re-mapping spaces that have been explored before. This undoubtedly leads to a huge waste of resources. Exploring only a small portion of missing areas may require several times the time and energy, resulting in extremely high costs. However, indoor environmental structures usually have certain similarities, and the spatial structure of missing areas can be predicted based on the environmental structure of the explored spaces.

[0071] (ii) Using LaMa networks for map prediction of missing spaces. Leveraging LaMa's powerful image perception capabilities, it can extract information from the entire map, not just local areas, which is very helpful for handling complex map layouts. However, networks with large receptive fields inevitably suffer from large parameter counts, high computational cost, and large model size, which are significant issues in edge scenarios with limited computing resources.

[0072] To address the aforementioned issues, this application provides a robot indoor exploration method based on a lightweight neural network prediction model. In this method, a local observation map of the target robot is acquired and input into a preset local prediction model to obtain a local prediction map. A global probability map is then updated based on the local prediction map. The updated global probability map is used to divide the area into free and uncertain regions. If boundary points exist for both free and uncertain regions, all boundary points are clustered. The distance from the target robot to the centroid of each cluster is calculated, and a target point is determined based on this distance. An exploration path is then determined based on the target point. After the robot reaches the target point, the steps of acquiring the local prediction map and determining a new target point are repeated until a global map is constructed. This solves the problems of poor performance, excessive computational burden, and low exploration efficiency in autonomous robot exploration in complex environments. This application combines unobserved region prediction with a real-time exploration strategy, reducing computational burden and improving autonomous exploration performance.

[0073] Specifically, Figure 4 This is a flowchart illustrating a robot indoor exploration method based on a lightweight neural network prediction model, provided in an embodiment of this application.

[0074] like Figure 4 As shown, this robot indoor exploration method based on a lightweight neural network prediction model includes the following steps:

[0075] In step S101, a local observation map of the target robot is obtained.

[0076] In step S102, the local observation map is input into the preset local prediction model to obtain the local prediction map, and the global probability map is updated according to the local prediction map. The free region and the uncertain region are divided according to the updated global probability map.

[0077] Optionally, in some embodiments, before inputting the local observation map into a preset local prediction model to obtain a local prediction map, the method includes: acquiring a local map prediction dataset; using the local map prediction dataset to train a neural network model to obtain a trained neural network model; and using a preset loss function to optimize the trained neural network model to obtain a preset local prediction model.

[0078] Specifically, embodiments of this application define a local observation map at time t. The map is set at robot position P. t Centered on the map, the perception range is L. The local map is constructed by defining a region within the entire environment map M that is twice the perception range, L.

[0079] LiDAR sensors can provide accurate observations within their sensing range, but they are limited by coverage area and occlusion effects caused by obstacles. Therefore, by setting the side length of the local map to 2L, the embodiments of this application can provide a certain prediction space for the model without significantly increasing the input size.

[0080] In obtaining local observation maps Subsequently, in this embodiment, codes are assigned to the free region, uncertain region, and obstacle region, and these codes are mapped to corresponding model features. These inputs are used for model training. Furthermore, during dataset construction, the pruning method described below is applied to extract each local observation map, and then the corresponding ground truth labels are annotated.

[0081]

[0082] in, Here, is a local observation map at time t, M is the global map, a is the horizontal coordinate of the two-dimensional map, and b is the vertical coordinate of the two-dimensional map. Let be the x-coordinate of robot P at time t, be the center coordinate of the local observation map (2), and L be the perception range. Let t be the ordinate of robot P.

[0083] This application modifies the simulation environment proposed by HouseExpo to collect local observation maps and ground truth labels. Based on the modified simulation environment, a local map prediction dataset, SenseMapDataset, was created using the KTH dataset. This dataset will be used for subsequent training and comparative experiments with our network model.

[0084] like Figure 5 As shown, this application embodiment constructs a lightweight neural network model for deployment on computationally limited robotic devices. Based on the UNet+Transformer framework, it incorporates the principles of VisionTransformer (ViT), seamlessly integrating the Transformer design into image processing, thus obtaining SenseMapNet. Its overall architecture is specifically tailored for local map prediction.

[0085] The input local observation map passes through two parallel processing streams: a convolutional encoder-decoder network and a Transformer-based encoding pipeline. The convolutional encoder extracts hierarchical spatial features, progressively reducing resolution and increasing feature depth. Simultaneously, the input map is divided into non-overlapping blocks, which are flattened and projected into the embedding space, then processed by the Transformer encoder. The Transformer module captures long-range dependencies and global spatial relationships. Multi-scale features extracted from the two streams are fused through skip connections and multi-resolution aggregation before the decoding process reconstructs the local prediction map. This dual-branch structure enables the model to simultaneously utilize fine-grained local spatial details and high-level contextual information, thereby improving the prediction accuracy and robustness of autonomous navigation tasks.

[0086] The UNet network demonstrates strong capabilities in local feature extraction and can recover spatial information through skip connections. On the other hand, the Transformer model excels in context understanding and global awareness. By combining the output of the Transformer encoder with skip connections, map prediction performance can be further improved.

[0087] The processing flow of the neural network model in this embodiment includes: taking 3×H×W map data as input, where the three channels correspond to the free region, uncertain region, and obstacle region, respectively. After the map is input into the model, the data is processed in units of patches. These patches undergo linear projection through flattened patches, and are then input into the Transformer encoder and convolutional encoder in the form of flattened patches, while preserving their original image format. The output of the Transformer encoder is divided into four equal parts, passed through different deconvolutional layers, and enlarged to different shapes. Then, these outputs are concatenated with the jumps of the convolutional encoder to provide global perception information for convolutional layers of different scales. After passing through multiple convolutional decoders, the output is an H×W×1 obstacle prediction map, which can predict the local map. This model can adjust the input and output channels of the convolutional layers and the embedding dimension of the Transformer encoder by modifying the base values. This flexibility simplifies the adjustment of the model size.

[0088] In addition, in this embodiment, the loss function can reflect the model's ability to predict unknown areas in a local observation map. The goal of the regression loss function in related technologies is to achieve a pixel-level perfect match between the model and the real label. However, in local map prediction tasks, free areas usually account for a larger proportion of the labels than obstacle areas. Pursuing a pixel-level perfect match may cause the model to become "conservative". Such models are usually overly similar to the observation map, thereby limiting the model's ability to effectively predict unknown areas.

[0089]

[0090] in, The mean squared error (MSE) loss is denoted by N, where N is the total number of map pixels.

[0091] To address this issue and enhance the model's predictive ability in unknown regions, embodiments of this application utilize feature reconstruction loss during training. The feature reconstruction loss uses a pre-trained loss network. To evaluate the feature reconstruction error between the predicted image and the target image, this strategy encourages the model's predictions to have features similar to the structural features computed by the loss network. Loss Network It is a deep convolutional neural network with L layers. Let... This represents the activation value of the j-th layer of the loss network when processing map m; if j is a convolutional layer, It will be a shape of C j ×H j ×W j The feature map. The feature reconstruction loss is defined as the Euclidean distance between the feature representations.

[0092]

[0093] in, The loss is for feature reconstruction, where L is the number of layers in the loss network, and C is the number of layers. j H represents the number of channels in the feature map of layer j. j W represents the height of the feature map at layer j. j The width of the feature map at layer j. To predict the activation value of the map in the j-th layer of the loss network (or, to predict the feature map of the map in the j-th layer of the loss network), When processing map m, the activation values ​​of the j-th layer of the loss network are lost.

[0094] The mean squared error (MSE) loss and feature reconstruction loss are respectively related to their corresponding weights w mse and w feat Multiply them, then sum them to get the final loss function.

[0095]

[0096] in, Let m be the predicted map and w be the actual map. mse The mean squared error loss weight value is... For mean squared error loss, w feat The weight values ​​for feature reconstruction loss. The feature reconstruction loss is used.

[0097] In this embodiment of the application, the loss network It is a 16-layer VGG network, pre-trained on ImageNet. mse =10, w feat =1.

[0098] Map exploration algorithms in related technologies typically rely on maps of known and unknown areas. In this embodiment, SenseMap utilizes the model's predictive ability for unknown areas to maintain a global probabilistic map. This makes the exploration algorithm dependent on a global probability map.

[0099] In actual execution, the global probability map is initialized. Its value is 0.5. At time t, the map is based on the robot's position P. t Updated, using local prediction map And update the weight α. Let λ be the local predicted map. Given the side length, we have:

[0100]

[0101] in, Let be the global probability map at time t. Let be the x-coordinate of any location on the global probability map at time t. Let be the ordinate of any location in the global probability map at time t, and α be the update weight value for updating the global probability map from the local prediction map. Let be the local predicted map at time t, where i is the x-coordinate of the predicted map and j is the y-coordinate of the predicted map. This is the global probability map at time t-1.

[0102] in:

[0103]

[0104] In step S103, it is determined whether there are boundary points of free regions and / or boundary points of uncertain regions. If there are boundary points of free regions and / or boundary points of uncertain regions, all boundary points are clustered, the distance between the target robot and the centroid of each cluster is calculated, the target point is determined based on the distance between the target robot and the centroid of each cluster, and the exploration path is determined based on the target point. After the robot reaches the target point, the steps of acquiring the local prediction map and determining the new target point are re-executed until the global map is constructed.

[0105] Furthermore, in some embodiments, dividing free regions and uncertain regions according to the updated global probability map includes: classifying regions in the global probability map that are less than a first preset threshold as free regions; classifying regions in the global probability map that are greater than a second preset threshold as obstacle regions, wherein the second preset threshold is greater than the first preset threshold; and classifying regions in the global probability map that are greater than or equal to the first preset threshold and less than or equal to the second preset threshold as uncertain regions.

[0106] Specifically, in combination Figure 6 As shown, where, Figure 6 The code in the text means:

[0107] Algorithm 1: SenseMap: Target Acquisition.

[0108] Input: Global probability map, robot state.

[0109] 1. Extract the set of boundary points F from the global probability map;

[0110] 2. If the set of boundary points F is not empty, then perform the following operations;

[0111] 3. Cluster the boundary points to obtain a cluster set C;

[0112] 4. For each cluster C in the cluster set C;

[0113] 5. Calculate the cost of this clustering based on the global probability map;

[0114] 6. Set the centroid of the cluster with the lowest cost as the path point;

[0115] 7. If the set of boundary points F is empty, the map building process is complete;

[0116] Output: Waypoints, map status (whether completed).

[0117] Obtaining the global probability map Then, the global probability map will be... All probabilities less than a threshold τ are classified as free regions, probabilities greater than a threshold v are classified as obstacle regions, and probabilities between τ and v are classified as uncertain regions. This embodiment uses a breadth-first search (BFS) algorithm to identify the boundary points of free and uncertain regions. If no boundary points are found, the exploration is considered complete. If boundary points exist, a connectivity algorithm is used to cluster all boundary points, and the centroid distance from the robot to each cluster is calculated, denoted as C.dis. The cost of each cluster is calculated by combining the probability value, distance, and its corresponding weight w. PROB and w dis The target point G is determined as shown in the formula below.i The centroid of the cluster that is set to minimize cost.

[0118]

[0119] Where C.cost is the total cost of a certain boundary cluster, used to measure the priority of that cluster as an exploration target, w PROB For uncertain weights, Let w be the probability value at coordinates (a, b) on the global probability map at time t. dis For distance weights.

[0120] In the algorithm of this application embodiment, the update weight α is set to 0.25. Since the number of free pixels in the training data significantly exceeds the number of obstacle pixels, the model is biased in predicting unknown regions as free space. To mitigate this bias and improve prediction accuracy, we set the thresholds to τ = 0.1 and v = 0.5. After obtaining the target point G... i Then, use A * Algorithm for obtaining path t And update robot status S t These steps will be repeated, such as... Figure 7 As shown, until the complete map is reconstructed, in which, Figure 7 The code in the text means:

[0121] Algorithm 2: SenseMap: Map Reconstruction.

[0122] Input: Initial state, maximum time T.

[0123] 1. Initialize the global probability map;

[0124] 2. For each time t;

[0125] 3. If a pathpoint exists and the robot's position at the previous moment is equal to the pathpoint, perform the following operations;

[0126] 4. Calculate the movement path: Based on the map from the previous moment, the robot's state from the previous moment, and the target point G, obtain the path;

[0127] 5. Update robot state: Based on the path and the robot state at the previous moment, obtain the current robot state;

[0128] 6. Update Map: Update the map based on the current robot status;

[0129] 7. Extract local observation map: Based on the current map and the current robot state, obtain the local observation map;

[0130] 8. Using SenseMapNet for prediction: Obtain a local prediction map based on the local observed map;

[0131] 9. Update the global probability map: Update the global probability map based on the local prediction map and the current robot state;

[0132] 10. Obtain the next path point: Based on the global probability map and the current robot state, obtain the next path point;

[0133] 11. If the path point does not exist;

[0134] 12. Exit the loop;

[0135] 13. Binarize the global probability map to obtain the final map: The final map M is derived from the global probability map at the last moment. Figure 2 Value-based;

[0136] Output: Final reconstructed map M.

[0137] In summary, this application introduces SenseMap, a method that uses neural network-based prediction to improve the efficiency of autonomous exploration in structured indoor environments. By integrating SenseMapNet (a lightweight hybrid architecture that combines a convolutional encoder and a Transformer-based perceptron), the robot is able to infer unobserved areas and optimize its exploration trajectory in real time.

[0138] Furthermore, through extensive experiments on the SenseMapDataset, this application demonstrates that the proposed method outperforms exploration strategies in related technologies, reducing exploration time while maintaining high reconstruction accuracy. Comparative evaluations show that SenseMapNet achieves excellent performance on multiple perceptual quality metrics (including SSIM, LPIPS, and FID) while maintaining a lightweight model suitable for airborne computing.

[0139] To verify the method of the embodiments of this application, some comparative experiments were conducted, as described below.

[0140] (1) Comparison of loss functions.

[0141] Class imbalance in the training data (where the free region is significantly larger than the obstacle) can lead to over-conservatism, such as... Figure 8As shown, the first two columns display the ground truth label map and the corresponding observation map. The remaining columns show the prediction maps generated by different models using mean squared error (MSE) loss or mixed loss (MIX) (i.e., combining perceptual loss and (MSE) loss). When using only mean squared error (MSE) loss, the prediction results are slightly lower. To mitigate this geometric distortion while maintaining reconstruction fidelity, a dual-branch architecture is used to synergistically combine perceptual loss with traditional MSE. This hybrid approach, through complementary loss components, simultaneously ensures pixel-level accuracy and structural consistency. The embodiments of this application use three metrics to evaluate performance: SSIM, LPIPS, and FID, such as... Figure 9 As shown, the SenseMapNet model demonstrates significant improvements in SSIM, LPIPS, and FID metrics. Furthermore, the addition of hybrid loss also significantly improves the SSIM and LPIPS scores of the UNet model, indicating the effectiveness of hybrid loss training in this task.

[0142] (2) Model performance comparison.

[0143] To compare the performance of models at different scales, embodiments of this application use a hybrid mean squared error loss (MSE) and perceptual loss to compare the number of parameters, SSIM, LPIPS, and FID among UNet, SenseMapNet, and SenseMapNetLarge. The original training protocol of LaMa-Fourier was also used to train the model on SenseMapDataset as a benchmark comparison.

[0144] like Figure 10 As shown, the first two columns represent the ground truth map and the corresponding observed map, while the remaining columns display the predicted maps generated by various models, including SenseMapNet, SenseMapNetLarge, UNet, and LaMa-Fourier. These visualizations demonstrate the effectiveness of different architectures in predicting local maps. Notably, SenseMapNet and its larger variants exhibit superior structure preservation and spatial consistency compared to UNet and LaMa-Fourier.

[0145] like Figure 11 As shown, SenseMapNet achieves outstanding performance while maintaining a lightweight model architecture. Compared to UNet, it improves in all four key metrics: parameter efficiency, SSIM, LPIPS, and FID. Furthermore, compared to LaMa, which utilizes adversarial learning, SenseMapNet also achieves better results in three key metrics: parameter efficiency, SSIM, and FID.

[0146] (3) Comparison of reconstruction efficiency.

[0147] This application embodiment conducted experiments on 10 different maps, each map was explored 50 times, and each exploration started from a different unknown area. The model performance was evaluated using three evaluation metrics: exploration time, exploration coverage ρ, and reconstruction accuracy (RA). The formulas for exploration coverage ρ and reconstruction accuracy (RA) are as follows, where I (condition) is an indicator function, with a value of 1 when the condition is true and a value of 0 when the condition is false.

[0148]

[0149] Where ρ is the exploration coverage, M(a,b) is the position with coordinates (a,b) in the robot's mapping, and M... true (a,b) represents the location of coordinates (a,b) on the real map, I is an indicator function that takes the value 1 when the condition in parentheses is true and 0 when it is false, R is a real number, and RA is the reconstruction accuracy.

[0150] Reconstruction results of different exploration methods, such as Figure 12 As shown, the change of coverage ρ over time is explored as follows: Figure 13 As shown. In Figure 12 In the diagram, the first column shows the ground truth map, the second and third columns represent the planned and SLAM maps, respectively, with explored areas marked in blue. The fourth column shows the predicted maps generated by different methods, where the Frontier-based method did not produce any predictions. The last column displays the binarized map obtained from the prediction output. The SenseMapNet and SenseMapNetLarge models generate more complete and accurate predictions than the traditional Frontier-based method, demonstrating their effectiveness in autonomous exploration. Figure 13 The changes in exploration coverage ρ over time are shown under different exploration methods. The coverage growth rate of the SenseMapNet and SenseMapNetLarge models is significantly faster than that of boundary-based exploration methods, indicating an improvement in exploration efficiency. Additionally, as... Figure 14 As shown, the SenseMap method significantly reduces the time required for exploration and improves exploration efficiency.

[0151] Therefore, the embodiments of this application can perform multi-robot autonomous exploration and map reconstruction in environments with limited computing power, and perform map prediction for locally unknown spaces, so that the entire framework can achieve a balance between maintaining exploration efficiency and reducing computing requirements.

[0152] The indoor robot exploration method based on a lightweight neural network prediction model proposed in this application involves acquiring a local observation map of the target robot and inputting it into a preset local prediction model to obtain a local prediction map. A global probability map is then updated based on the local prediction map. The updated global probability map is used to divide the area into free and uncertain regions. If boundary points exist for both free and uncertain regions, all boundary points are clustered. The distance from the target robot to the centroid of each cluster is calculated, and a target point is determined based on this distance. An exploration path is then determined based on the target point. After the robot reaches the target point, the steps of acquiring the local prediction map and determining a new target point are repeated until a global map is constructed. This method solves the problems of poor performance, excessive computational burden, and low exploration efficiency in autonomous robot exploration in complex environments. This application combines unobserved region prediction with a real-time exploration strategy to reduce computational burden and improve autonomous exploration performance.

[0153] Next, referring to the accompanying drawings, a robotic indoor exploration device based on a lightweight neural network prediction model, according to an embodiment of this application, is described.

[0154] Figure 15 This is a block diagram of a robot indoor exploration device based on a lightweight neural network prediction model according to an embodiment of this application.

[0155] like Figure 15 As shown, the robot indoor exploration device 10 based on a lightweight neural network prediction model includes: an acquisition module 100, a prediction module 200, and an exploration module 300.

[0156] The acquisition module 100 is used to acquire a local observation map of the target robot.

[0157] The prediction module 200 is used to input the local observation map into a preset local prediction model to obtain a local prediction map, update the global probability map according to the local prediction map, and divide the free region and uncertain region according to the updated global probability map.

[0158] The exploration module 300 is used to determine whether there are boundary points of the free region and / or the uncertain region. If there are boundary points of the free region and / or the uncertain region, all boundary points are clustered, the distance between the target robot and the centroid of each cluster is calculated, and the target point is determined based on the distance between the target robot and the centroid of each cluster. The exploration path is determined based on the target point. After the target robot reaches the target point, the steps of acquiring the local prediction map and determining the new target point are re-executed until the global map is constructed.

[0159] Optionally, in some embodiments, the prediction module 200 includes: a first partitioning unit, a second partitioning unit, and a third partitioning unit.

[0160] The first division unit is used to divide the regions in the global probability map that are less than a first preset threshold into the free regions based on the global probability map.

[0161] The second division unit is used to divide the regions in the global probability map that are greater than a second preset threshold into obstacle regions, wherein the second preset threshold is greater than the first preset threshold.

[0162] The third division unit is used to divide the regions in the global probability map that are greater than or equal to the first preset threshold and less than or equal to the second preset threshold into the uncertain regions.

[0163] Optionally, in some embodiments, before inputting the local observation map into a preset local prediction model to obtain a local prediction map, the prediction module 200 includes: an acquisition unit and an optimization unit.

[0164] The acquisition unit is used to acquire the local map prediction dataset.

[0165] The optimization unit is used to train the neural network model using the local map prediction dataset to obtain the trained neural network model, and to optimize the trained neural network model using a preset loss function to obtain the preset local prediction model.

[0166] Optionally, in some embodiments, the preset loss function is:

[0167]

[0168] in, Let m be the predicted map and w be the actual map. mse The mean squared error loss weight value is... For mean squared error loss, w feat The weight values ​​for feature reconstruction loss. The feature reconstruction loss is used.

[0169] Optionally, in some embodiments, updating the global probability map based on the local prediction map includes:

[0170] Based on a preset global probability map update calculation formula, the global probability map is updated according to the local prediction map. The preset global probability map update calculation formula is as follows:

[0171]

[0172] in, Let be the global probability map at time t. Let be the x-coordinate of any location on the global probability map at time t. Let be the ordinate of any location in the global probability map at time t, and α be the update weight value for updating the global probability map from the local prediction map. Let be the local predicted map at time t, where i is the x-coordinate of the predicted map and j is the y-coordinate of the predicted map. This is the global probability map at time t-1.

[0173] It should be noted that the foregoing explanation of the robot indoor exploration method based on a lightweight neural network prediction model also applies to the robot indoor exploration device based on a lightweight neural network prediction model in this embodiment, and will not be repeated here.

[0174] The indoor robot exploration device based on a lightweight neural network prediction model proposed in this application acquires a local observation map of the target robot and inputs it into a preset local prediction model to obtain a local prediction map. The global probability map is then updated based on the local prediction map. The updated global probability map is used to divide the area into free and uncertain regions. If boundary points exist in the free and / or uncertain regions, all boundary points are clustered. The distance from the target robot to the centroid of each cluster is calculated, and a target point is determined based on this distance. An exploration path is then determined based on the target point. After the robot reaches the target point, the steps of acquiring the local prediction map and determining a new target point are repeated until a global map is constructed. This solves the problems of poor performance, excessive computational burden, and low exploration efficiency in autonomous robot exploration in complex environments. This application combines unobserved region prediction with a real-time exploration strategy to reduce computational burden and improve autonomous exploration performance.

[0175] Figure 16 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include:

[0176] The memory 1601, the processor 1602, and the computer program stored on the memory 1601 and executable on the processor 1602.

[0177] When the processor 1602 executes the program, it implements the robot indoor exploration method based on a lightweight neural network prediction model provided in the above embodiments.

[0178] Furthermore, electronic devices also include:

[0179] Communication interface 1603 is used for communication between memory 1601 and processor 1602.

[0180] The memory 1601 is used to store computer programs that can run on the processor 1602.

[0181] The memory 1601 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0182] If the memory 1601, processor 1602, and communication interface 1603 are implemented independently, then the communication interface 1603, memory 1601, and processor 1602 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 16 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0183] Optionally, in a specific implementation, if the memory 1601, processor 1602, and communication interface 1603 are integrated on a single chip, then the memory 1601, processor 1602, and communication interface 1603 can communicate with each other through an internal interface.

[0184] The processor 1602 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0185] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described robot indoor exploration method based on a lightweight neural network prediction model.

[0186] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0187] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0188] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0189] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0190] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0191] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for indoor robot exploration based on a lightweight neural network prediction model, characterized in that, Includes the following steps: Obtain a local observation map of the target robot; The local observation map is input into a preset local prediction model to obtain a local prediction map, and the global probability map is updated according to the local prediction map. The free region and the uncertain region are divided according to the updated global probability map. Determine whether there are boundary points of the free region and / or the uncertain region. If there are boundary points of the free region and / or the uncertain region, cluster all boundary points, calculate the distance between the target robot and the centroid of each cluster, determine the target point based on the distance between the target robot and the centroid of each cluster, and determine the exploration path based on the target point. After the robot reaches the target point, re-execute the steps of acquiring the local prediction map and determining the new target point until the global map is constructed.

2. The method according to claim 1, characterized in that, The step of dividing the free region and the uncertain region according to the updated global probability map includes: Based on the global probability map, regions in the global probability map that are less than a first preset threshold are classified as the free regions; The regions in the global probability map that are greater than a second preset threshold are classified as obstacle regions, wherein the second preset threshold is greater than the first preset threshold; The regions in the global probability map that are greater than or equal to the first preset threshold and less than or equal to the second preset threshold are classified as the uncertain regions.

3. The method according to claim 1, characterized in that, Before inputting the local observation map into a preset local prediction model to obtain a local prediction map, the process includes: Obtain the local map prediction dataset; The neural network model is trained using the local map prediction dataset to obtain a trained neural network model, and the trained neural network model is optimized using a preset loss function to obtain the preset local prediction model.

4. The method according to claim 3, characterized in that, The preset loss function is: in, Let m be the predicted map and w be the actual map. mse The mean squared error loss weight value is... For mean squared error loss, w feat The weight values ​​for feature reconstruction loss. The feature reconstruction loss is used.

5. The method according to claim 1, characterized in that, The step of updating the global probability map based on the local prediction map includes: Based on a preset global probability map update calculation formula, the global probability map is updated according to the local prediction map. The preset global probability map update calculation formula is as follows: in, Let be the global probability map at time t. Let be the x-coordinate of any location on the global probability map at time t. Let be the ordinate of any location in the global probability map at time t, and α be the update weight value for updating the global probability map from the local prediction map. Let be the local predicted map at time t, where i is the x-coordinate of the predicted map and j is the y-coordinate of the predicted map. This is the global probability map at time t-1.

6. A robotic indoor exploration device based on a lightweight neural network prediction model, characterized in that, Includes the following steps: The acquisition module is used to acquire a local observation map of the target robot; The prediction module is used to input the local observation map into a preset local prediction model to obtain a local prediction map, update the global probability map according to the local prediction map, and divide the free region and uncertain region according to the updated global probability map. The exploration module is used to determine whether there are boundary points of the free region and / or the uncertain region. If there are boundary points of the free region and / or the uncertain region, all boundary points are clustered, the distance between the target robot and the centroid of each cluster is calculated, and the target point is determined based on the distance between the target robot and the centroid of each cluster. The exploration path is determined based on the target point. After the target robot reaches the target point, the steps of acquiring the local prediction map and determining the new target point are re-executed until the global map is constructed.

7. The apparatus according to claim 6, characterized in that, The prediction module includes: The first division unit is used to divide the regions in the global probability map that are less than a first preset threshold into the free regions based on the global probability map. The second division unit is used to divide the area in the global probability map that is greater than a second preset threshold into an obstacle area, wherein the second preset threshold is greater than the first preset threshold. The third division unit is used to divide the regions in the global probability map that are greater than or equal to the first preset threshold and less than or equal to the second preset threshold into the uncertain regions.

8. The apparatus according to claim 6, characterized in that, Before inputting the local observation map into a preset local prediction model to obtain a local prediction map, the prediction module includes: The acquisition unit is used to acquire the local map prediction dataset; The optimization unit is used to train the neural network model using the local map prediction dataset to obtain the trained neural network model, and to optimize the trained neural network model using a preset loss function to obtain the preset local prediction model.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, the processor executing the program to implement the robotic indoor exploration method based on a lightweight neural network prediction model as described in any one of claims 1-5.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the robotic indoor exploration method based on a lightweight neural network prediction model as described in any one of claims 1-5.

Images