Mobile robot path planning method and device based on entropy-aware feedback sampling

Abstract

The application discloses a mobile robot path planning method and device based on entropy perception feedback sampling, which first utilizes a characteristic linear modulation mechanism to explicitly encode obstacle constraints and passable structure information in a characteristic space, realizes deep fusion of environment semantics and geometric characteristics, secondly proposes an entropy perception feedback sampling strategy, realizes real-time monitoring of the uncertainty of network prediction by introducing global entropy evaluation, and realizes dynamic adjustment of sampling guide probability by combining a local collision feedback to build a punishment mechanism, finally integrates the prediction model and the sampling strategy into a traditional sampling planning algorithm as an enhanced module, and successfully deploys the module on a real mobile robot platform, and compared with a traditional sampling path planning algorithm, the application has significant optimization in path length, average search time and the number of expanded nodes in various simulation complex scenes and real scenes, and effectively improves path smoothness, and enhances the planning efficiency and robustness of the robot in an unknown environment.

Description

Technical Field

[0001] This invention relates to the field of mobile robot path planning technology, and in particular to a method and apparatus for mobile robot path planning based on entropy-aware feedback sampling. Background Technology

[0002] With the rapid development of mobile robot technology, it has been widely applied in search and rescue, environmental monitoring, intelligent logistics, autonomous driving, and collaborative inspection of unmanned systems. In these complex tasks, path planning, as the core of autonomous navigation systems, essentially seeks a collision-free and cost-optimal trajectory for the robot from its initial state to its target state within a complex configuration space. As operational scenarios shift from simple structured environments to highly unstructured and unknown dynamic environments, more stringent requirements are placed on path planning algorithms to achieve a high degree of balance between computational efficiency, theoretical completeness, and path optimality.

[0003] Currently, path planning methods for mobile robots mainly include classical algorithms and learning-based neural planning algorithms. Among classical algorithms, graph search algorithms ensure completeness through discretization, but face the curse of dimensionality in high-dimensional spaces. While artificial potential field methods are computationally efficient, their inherent local minima often cause robots to stagnate in complex obstacle groups. In contrast, sampling algorithms, represented by Fast Random Tree Search (RRT*), have become mainstream due to their computational advantages in high-dimensional spaces and asymptotic optimality. However, their uniform sampling strategy exhibits severe blindness when faced with complex geometric constraints, generating a large number of invalid nodes far from the optimal path, resulting in low sampling efficiency and slow convergence.

[0004] To address the blindness of sampling, learning-based neural planners have emerged. These methods utilize convolutional neural networks to extract map features and guide the search process to focus on promising regions by learning a non-uniform sampling distribution, thus significantly accelerating the convergence of the initial path. However, existing neural planning research still faces significant challenges in practical applications. First, the model prediction process often ignores the cognitive uncertainty inherent in the model itself, lacking effective confidence assessment methods when facing unfamiliar environments or complex topologies. Second, existing neural sampling guidance processes typically employ fixed probability distributions, lacking real-time feedback mechanisms. When the neural network prediction deviates from the optimal region or the robot gets trapped in a local optima, the algorithm cannot perceive the prediction failure and dynamically adjust its search strategy, making it highly susceptible to getting trapped in local optima in unfamiliar or highly dynamic environments, thus limiting its generalization ability and robustness.

[0005] To address the aforementioned issues, there is an urgent need for a mobile robot path planning method and device based on entropy-sensing feedback sampling. Summary of the Invention

[0006] The purpose of this invention is to address the aforementioned problems by providing a mobile robot path planning method and apparatus based on entropy-aware feedback sampling. This method introduces a feature linear modulation mechanism to enhance the feature encoding of obstacle constraints and passable structures, and combines this with an entropy-aware feedback sampling strategy to adaptively adjust the guidance weights and random exploration ratio of the neural network in real time. Furthermore, it effectively alleviates the adaptive deficiency problem caused by prediction uncertainty in neural planners, achieving efficient, robust, and asymptotically optimal path planning in complex and unknown environments. This provides core technical support for fully autonomous intelligent navigation of mobile robots and solves the technical problems of existing neural planners in complex, unstructured, or unknown environments, such as adaptive deficiency due to fixed guidance probabilities, susceptibility to local optima, and insufficient environmental constraint feature encoding.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: According to one aspect of the present invention, a method for path planning of a mobile robot based on entropy-aware feedback sampling is provided, comprising the following steps: S01. Construct an environmental map of the mobile robot's work scenario and the corresponding optimal path data, and create a path planning dataset for neural network training. S02. Construct a neural prior path prediction model based on the feature linear modulation mechanism, and train the model using the dataset to obtain the trained path prediction model. S03. An entropy-aware feedback sampling strategy is proposed. During the algorithm execution process, the global entropy of the output probability distribution of the path prediction model is calculated to quantify the confidence of the network prediction. The weight ratio of neural network guided sampling and random exploration sampling is dynamically adjusted in combination with local collision feedback. S04. The entropy-aware feedback sampling strategy is integrated with the neural prior path prediction model based on the feature linear modulation mechanism. Node expansion and search are performed in the constructed test set scenario to verify and output a collision-free planning path that meets environmental constraints. S5. Deploy the validated path planning method into the mobile robot, and plan and output the real motion control trajectory of the mobile robot online based on the real-time complex environment information obtained by the sensors.

[0008] Preferably, in step S01, the method for constructing the dataset includes the following steps: A random map generation algorithm is used to generate raster maps with different obstacle densities and topologies. On each map, start and target points are randomly sampled, and the corresponding optimal path is generated using a global path planning algorithm as the truth label. The map image, the coordinates of the start and end points, and the corresponding optimal path distribution map are labeled in pairs to form training sample pairs.

[0009] Preferably, in step S02, the path prediction model includes a map encoder, a path encoder, and a multi-scale decoder; The map encoder is used to extract the spatial topological features of obstacles in the environment and generate a multi-scale map feature set; The path encoder is used to predict feasible path trends based on environmental geometry under known start and end constraints, and generate multi-scale path features. The multi-scale decoder employs a symmetrical upsampling structure to deeply fuse the encoder's multi-scale path features, map features, and upsampled decoding features through skip connections.

[0010] Preferably, the path prediction model training method includes the following steps: The map features and the start and end point modulation signals are input into the path prediction model. The deviation between the predicted distribution and the true path value is calculated using the loss function. The network parameters are iteratively optimized by combining the backpropagation algorithm to establish a mapping relationship from environmental spatial constraints to the optimal path probability distribution.

[0011] Preferably, in step S03, the sampling method of the entropy-aware feedback sampling strategy includes the following steps: First, the Shannon entropy of the probability distribution graph of the neural network output is calculated as a measure of global prediction uncertainty; If the global entropy is lower than the preset threshold and the current sampling point does not trigger local collision feedback, then increase the sampling weight guided by the prior of the neural network. If the global entropy exceeds a preset threshold or a local collision feedback penalty is triggered, the prior weights are reduced and the proportion of random exploration sampling is increased simultaneously.

[0012] Preferably, in step S04, the method for generating a collision-free planned path includes the following steps: The trained neural prior model is used as an enhancement module and integrated into the sampling-based path planning algorithm; In different complex scenarios, the neural prior model guides node expansion, and the optimized path result is determined based on path length, search time, number of expanded nodes, and path curvature index.

[0013] Preferably, in step S05, the online planning method includes the following steps: Local map information is acquired in real time on a mobile robot platform using environmental perception sensors. The real-time map is input into the path prediction model to generate an online neural prior distribution, and dynamic sampling planning is performed in combination with real-time collision detection feedback. Based on the planned path points, the robot completes autonomous navigation through trajectory smoothing and underlying motion controller.

[0014] Preferably, the mobile robot path planning device based on entropy-aware feedback sampling includes: The environmental perception module is used to scan the surrounding environment of the mobile robot in real time, acquire environmental point cloud or depth data, and build a local real-time cost map for navigation. An edge computing module, which is communicatively connected to the environment perception module, is used to receive local map and target point information, generate online neural prior distribution in real time, calculate global entropy and combine it with the collision feedback of the environment perception module to execute a dynamic sampling strategy and plan a collision-free obstacle avoidance path online. The motion control and execution module is communicatively connected to the edge computing unit. It receives the planned path points and converts them into specific motor drive commands through trajectory smoothing processing to drive the mobile robot to complete autonomous navigation tasks in dynamic environments.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention first constructs diverse environmental maps and uses algorithms combined with thermal diffusion to generate continuous optimal path labels, providing a high-quality dataset for network training. Second, it constructs a neural prior path prediction model based on a feature linear modulation mechanism. By learning scaling and offset factors, it explicitly encodes environmental obstacle constraints and traversable topology in the feature space, achieving a high degree of integration between environmental semantics and planning tasks. Simultaneously, it proposes an entropy-aware feedback sampling strategy, which quantifies the global initial confidence by calculating the Shannon entropy of the network prediction probability graph and designs a dynamic penalty mechanism based on local collision feedback during sampling expansion, adaptively adjusting the weight ratio of neural network guidance and random exploration in real time. Finally, it integrates the prediction model and sampling strategy as plug-and-play enhancement modules into traditional sampling planning algorithms and successfully deploys them on a real mobile robot platform, achieving online obstacle avoidance and autonomous navigation.

[0016] 2. This invention innovatively combines an environmental feature linear modulation mechanism with a dynamic sampling guidance strategy based on global entropy and local feedback. This effectively overcomes the shortcomings of existing neural planners that are prone to getting trapped in local optima due to fixed guidance probabilities. While ensuring the asymptotic optimality of the path, it significantly improves the search efficiency and adaptability of the algorithm in unknown and highly unstructured environments, providing a new, efficient, and robust solution for fully autonomous intelligent navigation of mobile robots. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the present invention.

[0018] Figure 2This is a schematic diagram of the architecture of the neural prior path prediction model based on the feature linear modulation mechanism of the present invention.

[0019] Figure 3 This is a flowchart of the perception feedback sampling process of the present invention.

[0020] Figure 4 This is a grouped bar chart comparing the key performance indicators of four path planning algorithms for the perceptual feedback neural prior method proposed in this invention.

[0021] Figure 5 This is a visual diagram illustrating the optimization effect of the perceived feedback neural prior method of the present invention in an irregular obstacle scene.

[0022] Figure 6 This is a visual illustration of the optimization effect of the perceived feedback neural prior method of the present invention in a square obstacle scene.

[0023] Figure 7 This is a visual diagram illustrating the optimization effect of the perceived feedback neural prior method of the present invention in a narrow channel scenario.

[0024] Figure 8 This invention relates to a real-world experimental environment constructed using a mobile robot and a corresponding occupancy grid diagram.

[0025] Figure 9 This is a visual diagram illustrating the path planning results of different algorithms in a real-world scenario using the perceptual feedback neural prior method proposed in this invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and preferred embodiments. However, it should be noted that many details listed in the specification are merely to provide the reader with a thorough understanding of one or more aspects of the invention, and these aspects of the invention can be achieved even without these specific details.

[0027] Please see Figures 1 to 9 This invention provides a method and apparatus for mobile robot path planning based on entropy-aware feedback sampling, the technical solution of which is as follows: The mobile robot path planning method based on entropy-aware feedback sampling includes the following steps: S01. Construct an environmental map of the mobile robot's work scenario and the corresponding optimal path data, create a path planning dataset for neural network training, and divide the dataset.

[0028] Specifically, a random map generation algorithm was used to generate grid maps containing scenes with irregular and square obstacles as the training dataset. The map size was uniformly 100×100mm, with 6000 maps per category, totaling 12000 maps. To ensure the optimality and consistency of the paths learned by the neural network, this embodiment used the Theta* algorithm to sample the optimal paths from the 12000 random maps as label data. Theta* algorithm is an arbitrary-angle path planning algorithm based on the A* algorithm, capable of generating smoother and shorter paths while ensuring collision-free paths, which better meets the actual movement requirements of robots than the traditional A* algorithm. Simultaneously, this embodiment adopted an optimal path labeling method based on thermal diffusion. This method uses a path point as the center and expands spatially outwards to extend the path into a thermal region with a radius of 5, thereby constructing a continuous path probability distribution label. This approach imparts a certain degree of spatial tolerance to the path, allowing for reasonable deviations in the predicted path near the optimal path. Furthermore, to objectively evaluate the generalization ability of the proposed method in unknown and complex environments, this embodiment also constructed a narrow-channel scene map as an independent test set. This test set is strictly independent of the model training phase and does not involve any network weight updates. It is designed to verify the inference performance of the prediction network when faced with novel and highly challenging obstacle topologies.

[0029] S02. Construct a neural prior path prediction model based on the feature linear modulation mechanism, train the model using the dataset, explicitly encode obstacle constraints and passable structure information in the feature space, and obtain the trained path prediction model.

[0030] like Figure 2 As shown, the construction and training process of the neural prior path prediction model FiLMPath-Net includes the introduction of a Feature Linear Modulation (FiLM) mechanism, which allows map features to participate in the path feature generation process as conditional variables. FiLM provides an explicit conditional constraint encoding method, enabling the network to dynamically modulate environmental structure information in the feature space, thereby improving the expressive power of path distribution prediction.

[0031] The input to FiLMPath-Net consists of two branches: a map feature branch and a map feature branch. (Obstacle map) and mission objective branch (A binary graph including the start and end points). The network consists of three parts: a map encoder, a map encoder, and a map encoder. Path encoder And a multi-scale decoder. This includes the following detailed steps: S0201, Map Encoder It is responsible for extracting the spatial topological features of obstacles in the environment. Through four layers of consecutive convolution and downsampling operations, it generates a multi-scale map feature set. These features capture local obstacle edge information and contain global connectivity context, providing a geometric benchmark for subsequent path feature modulation. Their mathematical formula is as follows: in, Representing the Feature map of the layer; These are the input features for the previous layer; This is a two-dimensional convolution operation used to extract local spatial structure information; This is a batch normalization operation used to mitigate internal covariate bias and accelerate network convergence; It is a linear rectifier activation function used to introduce nonlinear expressive power and enhance the sparsity of features.

[0032] S0202, Path Encoder The core task is to predict feasible path trends based on environmental geometry, given the constraints of a known start and end point. This embodiment introduces the FiLM mechanism to inject map features as conditional parameters into the path features. For each layer of path features... This embodiment utilizes the corresponding map features. Generate a set of modulation parameters and The mathematical definition of the FiLM layer is as follows: in, Use the Sigmoid activation function to ensure the scaling factor. exist This serves to select features; This indicates multiplication by channel. The offset is determined by the characteristics of the open space guiding path.

[0033] Each encoding layer is nested within a residual block to ensure gradient flow and feature stability in deep networks. S0203, the decoding stage employs a symmetrical upsampling structure, using skip connections to deeply fuse the encoder's multi-scale path features and map features with the upsampled decoding features. Decoder... The input of a layer consists of three parts concatenated together: in, By stitching together the channels, the final generated guide map is ensured to have both the precise geometric boundaries of the underlying map and the global path connectivity of the high-level semantics.

[0034] Finally, a path probability distribution map with the same size as the original map is generated by using a 1×1 convolution and a sigmoid activation function on the output header. The pixel values ​​in this probability map represent the confidence that the point is on the optimal path, providing input for the subsequent entropy-aware feedback sampling strategy. S0204. To ensure consistency between the path probability distribution output by FiLMPath-Net and the true optimal path during training, this embodiment uses spatial cross-entropy loss as the optimization objective. (Path probability map output by the model) This represents the confidence score of each pixel belonging to the optimal path, with the true label being the corresponding path annotation map. Its loss function is defined as follows: in, Represents the probability prediction graph In position Pixel value at that location, These are the corresponding real tags.

[0035] By minimizing this loss function, the network can learn the conditional dependencies between map geometric constraints and path distribution, resulting in a predicted path probability map that outputs low confidence in obstacle areas and high confidence in feasible path areas, thus providing reliable probabilistic guidance for subsequent path planning. S03. An entropy-aware feedback sampling strategy is proposed. During the algorithm execution process, the global entropy of the output probability distribution of the path prediction model is calculated to quantify the confidence of the network prediction. The weight ratio of neural network guided sampling and random exploration sampling is dynamically adjusted in combination with local collision feedback.

[0036] like Figure 3 As shown, the dynamic adjustment process of the Entropy-aware Feedback Sampling (EFS) strategy consists of three parts: a mixed sampling distribution strategy, an entropy-based initial confidence assessment, and feedback-based runtime dynamic adjustment. Specifically, it includes the following steps: S0301. In the t-th iteration of RRT*, the sampled state follows a mixed distribution consisting of the network's predicted distribution and a globally uniform distribution. Its probability density function is defined as follows: in, Probability map corresponding to the output of FiLMPath-Net ; Uniform sampling of the entire state space. For the first The dynamic bias factor of each step determines the planner's level of trust in the neural network guidance.

[0037] S0302. To measure the global certainty of network predictions, this embodiment introduces Shannon entropy as a metric. Intuitively, when the prediction graph presents a clear, narrow path, the entropy value is low, indicating that the network has high prediction confidence; conversely, when the prediction graph is diffusely distributed, the entropy value is high, reflecting uncertainty in the network's path judgment.

[0038] set up For probability graphs The Middle Given the normalized probability values ​​of pixels, the entropy of the prediction distribution is defined as: Therefore, the initial confidence coefficient is defined. for: in, This represents the maximum entropy value under uniform distribution conditions. This definition ensures that as network predictions become more concentrated and path confidence increases, The larger the network, the more it can fully utilize the prior information provided by the network in the early stages of planning.

[0039] S0303. Relying solely on global entropy values ​​is insufficient to effectively address the problem of mispredictions in unseen maps by deep learning, i.e., the model provides high-confidence predictions for obstacles or inaccessible areas. Therefore, this embodiment further proposes a penalty factor based on runtime feedback. The network trust level is dynamically adjusted based on the actual effect of sampling expansion. Definition Deadline The runtime penalty factor is the number of consecutive sampling failures within a high-probability region of the network. Using exponential decay: in, Based on the scaling constant, The hyperparameter for controlling the decay rate of the penalty intensity.

[0040] When the planner appears consecutively in the high-probability region suggested by the network... When the extension fails It will drop rapidly, thereby forcing the bias factor to... Reducing the sampling size gradually degenerates the sampling strategy into the traditional uniform sampling RRT*, thus effectively escaping local optima caused by misguided approaches. Conversely, once the expansion of any region is successful, Reset to 0 Restored to initial level, network booting is now active again.

[0041] Combining the above global prior and local feedback information, the first dynamic bias of the step The child is defined as: in, This indicates that the value is restricted to a certain range. Within the interval. Through this adjustment mechanism, the EFS method can accelerate planning convergence when the network prediction is reliable, and automatically back to unbiased sampling when the network guidance fails, thereby improving efficiency while maintaining the probabilistic completeness of the algorithm.

[0042] S04. Integrate the entropy-aware feedback sampling strategy with the neural prior path prediction model based on feature linear modulation mechanism, perform node expansion and search in the test set scene, verify and output the collision-free planning path that meets the environmental constraints.

[0043] Specifically, the trained neural prior model is seamlessly integrated into RRT* and its various variants as a plug-and-play sampling enhancement module. Guided by the neural prior distribution, the algorithm prioritizes sampling in foreground regions with high probability and incorporates a perceptual feedback strategy to escape local optima in real time. In a test set scenario, comparative experiments were conducted to statistically analyze path length, average search time, number of expanded nodes, and path smoothness metrics, verifying the method's ability to improve search efficiency while adapting to various complex topological environments.

[0044] S05. Deploy the verified path planning method in the mobile robot device, and plan and output the real motion control trajectory of the mobile robot online based on the real-time complex environment information obtained by the sensor.

[0045] Specifically, an onboard computing unit is deployed on the mobile robot platform to acquire local environmental information in real time using LiDAR or depth cameras and construct a real-time cost map. The real-time map and target point coordinates are input into the model to generate an online neural prior distribution in real time. Combined with local collision detection feedback from the onboard sensors, an obstacle avoidance path is planned online using a dynamic sampling strategy. The generated path, after smoothing through cubic spline interpolation, is sent to the underlying motion controller to drive the mobile robot to complete autonomous navigation tasks in a dynamic environment.

[0046] like Figure 8As shown, the present invention also provides a mobile robot path planning device based on entropy perception feedback sampling, the device including an environment perception module, an edge computing module and a motion control and execution module.

[0047] The environmental perception module includes a lidar, which is used to scan the vehicle's surroundings in real time, acquire environmental point cloud or depth data, and build a local real-time cost map for navigation.

[0048] The edge computing module communicates with the environment perception module and internally carries a neural prior path prediction model based on feature linear modulation. The edge computing module receives local map and target point information, generates online neural prior distribution in real time, calculates global entropy, and combines collision feedback from the environment perception module to execute a dynamic sampling strategy, thereby planning a collision-free obstacle avoidance path online.

[0049] The motion control and execution module includes a low-level motion controller and a chassis drive mechanism, and is communicatively connected to the edge computing module. The motion control and execution module receives the planned path points, processes them through trajectory smoothing, and converts them into specific motor drive commands to drive the mobile robot to complete autonomous navigation tasks in dynamic environments.

[0050] To verify the effectiveness of the entropy-aware feedback neural prior method, in this embodiment, the entropy-aware feedback neural prior method is integrated into four basic path planning algorithms—RRT*, Q-RRT*, Informed-RRT, and Bi-RRT—for comparative evaluation. Quantitative results are shown in Table 1.

[0051] Table 1. Quantitative comparison of performance enhancements of different path planning algorithms To visually demonstrate the performance improvement, such as Figure 5As shown, the performance of each basic algorithm and the algorithm using the entropy-aware feedback neural prior method is visualized by grouping and comparing them in the form of a bar chart. It can be observed that after introducing the entropy-aware feedback neural prior method, all basic algorithms achieved significant and consistent improvements in various indicators across the three test scenarios. Regarding path quality, the path length generated by the algorithm using the entropy-aware feedback neural prior method was shortened by an average of approximately 9.2% compared to the original algorithm, and the path smoothness was improved. This indicates that the entropy-aware feedback neural prior method can effectively eliminate redundant paths, helping the algorithm find shorter and smoother trajectories that are easier for the robot's underlying control to execute. In terms of computational efficiency, the entropy-aware feedback neural prior method effectively guides the sampling space, reducing ineffective exploration by the algorithm. The average search time of each algorithm decreased by approximately 37.7%, and the number of expanded nodes decreased by approximately 41.4%, which accelerated the convergence process. Regarding robustness and generalization ability, the entropy-aware feedback neural prior method increased the success rate of each algorithm to 100% in scenarios 1 and 2. In scenario 3, the success rate of the traditional algorithm was between 66% and 72%, while the success rate was improved by 24% to 34% after introducing the entropy-aware feedback neural prior method. This demonstrates that the entropy-aware feedback neural prior method can enhance the adaptability of the original algorithm when facing unknown and complex scenarios.

[0052] To visually demonstrate the effectiveness of the entropy-aware feedback neural prior method, Figure 5 , 6 Comparing the sampling distribution and path planning results with EFN-RRT*, Q-RRT*, Informed-RRT*, and Bi-RRT in three scenarios, the sampling points and expansion trees almost cover the entire free space, exposing a large amount of blind search and redundant computation. However, in EFN-RRT*, EFN-Q-RRT*, EFN-Informed-RRT*, and EFN-Bi-RRT*, after introducing the entropy-aware feedback neural prior method, the sampling process is highly guided, focusing on the high-probability heatmap regions predicted by the network. For example... Figure 5 As shown, even after numerous iterations, the basic algorithm struggles to identify narrow gaps between obstacles, leading to severe path detours. Figure 6 As shown, the enhanced path of the entropy-aware feedback neural prior method is straighter and smoother, reducing the jagged fluctuations common in traditional RRT*-type algorithms. Specifically, comparing (c) and (g) reveals that EFN-Informed-RRT* eliminates the large-area sampling redundancy of Informed-RRT*, ensuring that sampling points are fully focused on the prior distribution. Finally, in Figure 7In the generalization test, the entropy-aware feedback neural prior method exhibits a strong directional guidance capability. Even when faced with an unseen map, it can still provide accurate predictive priors, ensuring that the algorithm still has excellent convergence speed in unknown environments.

[0053] To verify the effectiveness of the entropy-aware feedback neural prior method in real-world scenarios, this embodiment utilizes a tracked mobile robot to build a software testing environment based on Ubuntu 20.04 LTS and ROS Noetic. Figure 8 As shown, the platform incorporates an NVIDIA Jetson Orin Nano edge computing module to meet the computing power requirements for real-time processing of underlying sensor data and execution of target path planning. Simultaneously, the platform's perception system is equipped with a 10Hz, 25m ranging 360° LiDAR and a six-axis IMU to achieve environmental scanning and accurate attitude calculation. The experiment, based on the ROS framework, employs a point cloud algorithm, fusing / scan laser data and / odom odometry information to construct a 3D global point cloud map. Subsequently, a height threshold is set to perform pass-through filtering on the point cloud to remove ground and ceiling features, and the remaining obstacle features are projected along the Z-axis onto a plane, thus transforming it into a 2D occupancy grid map with a resolution of 0.05m / pixel, providing high-precision environmental prior information for subsequent planning tasks. In terms of software architecture, the proposed entropy-aware feedback neural prior method is integrated into the navigation system in the form of ROS nodes. The algorithm generates the optimal path based on the global map published by map_server, the robot's real-time pose, and the target point, and converts it into control commands that conform to the kinematic constraints of the chassis. Finally, it continuously sends speed commands through the / cmd_vel topic to achieve smooth autonomous obstacle avoidance and navigation in complex scenarios.

[0054] Based on the aforementioned platform, this embodiment comprehensively compares EFN-Bi-RRT* with RRT*, Bi-RRT*, and MPT-RRT* algorithms. To ensure objectivity, all algorithms were tested repeatedly 10 times under identical conditions, and the evaluation metrics are shown in Table 2.

[0055] Table 2 Performance Comparison of Path Planning Algorithms in Real-World Scenarios The EFN-Bi-RRT* algorithm demonstrates comprehensive performance advantages in real-world scenarios. Specifically, thanks to the heuristic guidance of the entropy-aware feedback neural prior method, EFN-Bi-RRT* significantly reduces blind sampling, achieving path length, average execution time, and number of sampling nodes of 14.47 meters, 0.68 seconds, and 1735, respectively, representing improvements of 10.62%, 35.85%, and 19.82% compared to Bi-RRT*. In terms of path quality, its smoothness reaches an optimal 41.53, better fitting the kinematic constraints of the robot chassis. Regarding stability, EFN-Bi-RRT* maintains a 100% planning success rate, improving upon RRT*, Bi-RRT*, and MPT-RRT* by 20%, 20%, and 10%, respectively. Figure 9 As shown, compared to the trajectories with redundant inflection points generated by the comparison algorithms, the path generated by EFN-Bi-RRT* is the smoothest, which fully demonstrates its efficiency and robustness in real-world scenarios.

[0056] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A mobile robot path planning method based on entropy-aware feedback sampling, characterized in that, Includes the following steps: S01. Construct an environmental map of the mobile robot's work scenario and the corresponding optimal path data, and create a path planning dataset for neural network training. S02. Construct a neural prior path prediction model based on the feature linear modulation mechanism, and train the model using the dataset to obtain the trained path prediction model. S03. An entropy-aware feedback sampling strategy is proposed. During the algorithm execution process, the global entropy of the output probability distribution of the path prediction model is calculated to quantify the confidence of the network prediction. The weight ratio of neural network guided sampling and random exploration sampling is dynamically adjusted in combination with local collision feedback. S04. The entropy-aware feedback sampling strategy is integrated with the neural prior path prediction model based on the feature linear modulation mechanism. Node expansion and search are performed in the constructed test set scenario to verify and output a collision-free planning path that meets environmental constraints. S05. Deploy the validated path planning method into the mobile robot, and plan and output the real motion control trajectory of the mobile robot online based on the real-time complex environment information obtained by the sensors.

2. The mobile robot path planning method based on entropy-aware feedback sampling according to claim 1, characterized in that: In step S01, the method for constructing the dataset includes the following steps: A random map generation algorithm is used to generate raster maps with different obstacle densities and topologies. On each map, start and target points are randomly sampled, and the corresponding optimal path is generated using a global path planning algorithm as the truth label. The map image, the coordinates of the start and end points, and the corresponding optimal path distribution map are labeled in pairs to form training sample pairs.

3. The mobile robot path planning method based on entropy-aware feedback sampling according to claim 1, characterized in that: In step S02, the path prediction model includes a map encoder, a path encoder, and a multi-scale decoder; The map encoder is used to extract the spatial topological features of obstacles in the environment and generate a multi-scale map feature set; The path encoder is used to predict feasible path trends based on environmental geometry under known start and end constraints, and generate multi-scale path features. The multi-scale decoder employs a symmetrical upsampling structure to deeply fuse the encoder's multi-scale path features, map features, and upsampled decoding features through skip connections.

4. The mobile robot path planning method based on entropy-aware feedback sampling according to claim 3, characterized in that: The path prediction model training method includes the following steps: The map features and the start and end point modulation signals are input into the path prediction model. The deviation between the predicted distribution and the true path value is calculated using the loss function. The network parameters are iteratively optimized by combining the backpropagation algorithm to establish a mapping relationship from environmental spatial constraints to the optimal path probability distribution.

5. The mobile robot path planning method based on entropy-aware feedback sampling according to claim 1, characterized in that: In step S03, the sampling method of the entropy-aware feedback sampling strategy includes the following steps: First, the Shannon entropy of the probability distribution graph of the neural network output is calculated as a measure of global prediction uncertainty; If the global entropy is lower than the preset threshold and the current sampling point does not trigger local collision feedback, then increase the sampling weight guided by the prior of the neural network. If the global entropy exceeds a preset threshold or a local collision feedback penalty is triggered, the prior weights are reduced and the proportion of random exploration sampling is increased simultaneously.

6. The mobile robot path planning method based on entropy-aware feedback sampling according to claim 1, characterized in that: In step S04, the method for generating a collision-free planned path includes the following steps: The trained neural prior model is used as an enhancement module and integrated into the sampling-based path planning algorithm; In different complex scenarios, the neural prior model guides node expansion, and the optimized path result is determined based on path length, search time, number of expanded nodes, and path curvature index.

7. The mobile robot path planning method based on entropy-aware feedback sampling according to claim 1, characterized in that: In step S05, the online planning method includes the following steps: Local map information is acquired in real time on a mobile robot platform using environmental perception sensors. The real-time map is input into the path prediction model to generate an online neural prior distribution, and dynamic sampling planning is performed in combination with real-time collision detection feedback. Based on the planned path points, the robot completes autonomous navigation through trajectory smoothing and underlying motion controller.

8. A mobile robot path planning device based on entropy-aware feedback sampling, used to execute the mobile robot path planning method based on entropy-aware feedback sampling as described in any one of claims 1-7, characterized in that, include: The environmental perception module is used to scan the surrounding environment of the mobile robot in real time, acquire environmental point cloud or depth data, and build a local real-time cost map for navigation. An edge computing module, which is communicatively connected to the environment perception module, is used to receive local map and target point information, generate online neural prior distribution in real time, calculate global entropy and combine it with the collision feedback of the environment perception module to execute a dynamic sampling strategy and plan a collision-free obstacle avoidance path online. The motion control and execution module is communicatively connected to the edge computing unit. It receives the planned path points and converts them into specific motor drive commands through trajectory smoothing to drive the mobile robot to complete autonomous navigation tasks in dynamic environments.

Images