Planet vehicle trafficability prediction method

Abstract

The invention provides a method for predicting the trafficability of a planet vehicle, and relates to the technical field of deep space exploration, and the method comprises the steps: obtaining a multi-modal topographic map of a target star catalogue region, the current state of the planet vehicle, and candidate motion primitives, and inputting a trafficability prediction model; performing feature extraction on the multi-modal topographic map through a feature extractor of the trafficability prediction model to obtain topographic geometry and mechanics coupled shared features; performing combined coding on the current state of the planet vehicle and the candidate motion primitives through a motion encoder of the model to obtain a combined coding result; performing dimension matching and splicing on the combined coding result and the shared features to obtain a one-dimensional fusion feature tensor; and through a loss predictor of the model, loss prediction is carried out according to the one-dimensional fusion feature tensor, and the movement cost of the planet vehicle during execution of the candidate motion primitives is obtained. By learning and fusing the multi-modal information and the vehicle state and motion information, the calculation process is simplified, and the prediction efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep space exploration, in particular to a planetary rover passability prediction method. BACKGROUND

[0002] With the continuous advancement of deep space exploration missions, planetary rovers have become the core mobile platform for extraterrestrial planet surface exploration. Due to the complex and variable surface environment of the planet, there are soft soil, rocks, sand dunes and steep slopes and other terrains, which pose a serious challenge to the safe movement of the planetary rover. Therefore, it is usually necessary to predict the passability of the planetary rover based on the geometric features of the terrain or the dynamics of the vehicle.

[0003] In related technologies, based on the geometric features of the terrain, since only parameters such as slope and relief are relied on, it is often difficult to accurately identify the terrain that is geometrically feasible but mechanically infeasible, and based on the dynamics of the vehicle, it relies on soil mechanical parameters that are difficult to obtain in real time, and the calculation is complex, the reasoning speed is slow, and it cannot be predicted in real time, which can easily lead to the planetary rover being trapped in a safety risk, and it is difficult to provide sufficient protection for the safe and efficient movement of the planetary rover in the complex surface environment of the planet. SUMMARY

[0004] The problem solved by the present application is how to improve the prediction effect of the passability of the planetary rover.

[0005] To solve the above problems, the present application provides a planetary rover passability prediction method, comprising: obtaining a multi-modal terrain map of a target planet surface area, a current state of a planetary rover and a candidate motion primitive; inputting the multi-modal terrain map, the current state and the candidate motion primitive into a passability prediction model; extracting features of the multi-modal terrain map through a feature extractor of the passability prediction model to obtain shared features of terrain geometry and mechanics coupling; combining and encoding the current state of the planetary rover and the candidate motion primitive through a motion encoder of the passability prediction model to obtain a combination encoding result; dimensionally matching and splicing the combination encoding result and the shared features to obtain a one-dimensional fusion feature tensor; predicting a loss according to the one-dimensional fusion feature tensor through a loss predictor of the passability prediction model to obtain a movement cost of the planetary rover when executing the candidate motion primitive.

[0006] Optionally, the obtaining of the multi-modal terrain map of the target planet surface area, the current state of the planetary rover and the candidate motion primitive comprises: acquire a local elevation map, a ground friction characteristic map and a ground softness characteristic map within a preset range of the star buggy, with the current position of the star buggy as the center; stack the local elevation map, the ground friction characteristic map and the ground softness characteristic map by channel to generate the multi-modal terrain map; acquire a horizontal coordinate, a vertical coordinate of the center of the star buggy in a preset coordinate system and a heading angle of the vehicle body longitudinal axis of the star buggy relative to the y-axis of the preset coordinate system; compose the current state according to the horizontal coordinate, the vertical coordinate and the heading angle; generate the candidate motion primitive according to a preset steering angle and a preset moving direction of the star buggy.

[0007] Optionally, the feature extractor of the passability prediction model extracts features from the multi-modal terrain map to obtain shared features of terrain geometry and mechanics, including: input the local elevation map, the ground friction characteristic map and the ground softness characteristic map in the multi-modal terrain map into multiple channels of the feature extractor respectively; extract features from the terrain data input by the multiple channels layer by layer through the convolution structure of the feature extraction layer to obtain geometric features in the local elevation map and mechanical property distribution features in the ground friction characteristic map and the ground softness characteristic map; couple the geometric features and the mechanical property distribution features across channels through the feature fusion module of the feature extraction layer to generate the shared features.

[0008] Optionally, the motion encoder of the passability prediction model combines the current state of the star buggy and the candidate motion primitive to obtain a combined encoding result, including: combine the horizontal coordinate, the vertical coordinate and the heading angle in the current state with the preset steering angle and the preset moving direction in the candidate motion primitive to form a motion-state joint parameter set; perform numerical mapping and dimension conversion on the motion-state joint parameter set through the fully connected layer of the motion encoder to obtain a feature vector of the motion-state joint parameter set; normalize the feature vector to obtain the combined encoding result.

[0009] Optionally, the combined encoding result and the shared features are dimensionally matched and spliced to obtain a one-dimensional fusion feature tensor, including: flatten the shared features to convert the feature matrix of the shared features into a one-dimensional feature vector; vector splicing the combination encoding result and the one-dimensional feature vector in a first preset order to obtain the one-dimensional fusion feature tensor.

[0010] Optionally, the loss predictor of the passability prediction model performs loss prediction according to the one-dimensional fusion feature tensor to obtain a movement cost of the star vehicle when executing the candidate motion primitive, including: inputting the one-dimensional fusion feature tensor into a shared fully connected layer of the loss predictor to extract a high-dimensional fusion feature; inputting the high-dimensional fusion feature into multiple parallel branches of the loss predictor respectively to perform loss prediction to obtain a predicted value corresponding to each parallel branch; generating the movement cost according to the predicted value of each parallel branch.

[0011] Optionally, the parallel branches include a time branch, an energy branch, and a danger degree branch, and the inputting the high-dimensional fusion feature into multiple parallel branches of the loss predictor respectively to perform loss prediction to obtain a predicted value corresponding to each parallel branch, including: mapping the high-dimensional fusion feature through a first sub fully connected layer of the time branch to obtain a predicted time cost required for executing the candidate motion primitive, and taking the predicted time cost as the predicted value of the time branch; mapping the high-dimensional fusion feature through a second sub fully connected layer of the energy branch to obtain a predicted energy cost required for executing the candidate motion primitive, and taking the predicted energy cost as the predicted value of the energy branch; mapping the high-dimensional fusion feature through a third sub fully connected layer of the danger degree branch to obtain a predicted danger degree value when executing the candidate motion primitive, and taking the predicted danger degree value as the predicted value of the danger degree branch.

[0012] Optionally, the vector splicing or weighted fusion according to the predicted value of each parallel branch to generate the movement cost, including: vector combining the predicted time cost, the predicted energy cost, and the predicted danger degree value corresponding to the time branch, the energy branch, and the danger degree branch respectively in a second preset order to obtain the movement cost; or, configuring corresponding weights for the predicted time cost, the predicted energy cost, and the predicted danger degree value respectively; weighted fusion according to the weights corresponding to the predicted time cost, the predicted energy cost, and the predicted danger degree value respectively to obtain the movement cost.

[0013] Optionally, the star vehicle passability prediction method further comprises: constructing the passability prediction model. The constructing the passability prediction model comprises: In a simulated star table environment, a plurality of modal terrain maps, current states, candidate motion primitives and actual time costs, energy costs and danger labels corresponding to the candidate motion primitives of the star vehicle under different terrains are collected to generate a training sample set; The training sample set is subjected to terrain map random rotation processing, height drift processing and noise disturbance processing to obtain an enhanced training sample set; The enhanced training sample set is input into an initial convolutional neural network, and the initial convolutional neural network is trained through a multi-task loss function; The initial convolutional neural network after training is used as the passability prediction model.

[0014] Optionally, the inputting the enhanced training sample set into the initial convolutional neural network and training the initial convolutional neural network through the multi-task loss function comprises: The plurality of modal terrain maps, current states and candidate motion primitives in the enhanced training sample set are used as inputs of the initial convolutional neural network, and the actual time cost, actual energy cost and actual danger label in the enhanced training sample set are used as supervision targets; A multi-task loss function is constructed, and the multi-task loss function is a weighted sum of a time cost loss, an energy cost loss and a danger degree loss, wherein the time cost loss is a mean square error between a predicted time cost and the actual time cost; the energy cost loss is a mean square error between a predicted energy cost and the actual energy cost; and the danger degree loss is a cross-entropy loss or a mean square error between a predicted danger degree value and the actual danger label; The parameters of the initial convolutional neural network are iteratively optimized through a back propagation algorithm, and the loss value determined by the multi-task loss function is used to determine when the training of the initial convolutional neural network is completed.

[0015] The passability prediction method of the star vehicle of the present application can provide richer basic data for subsequent feature extraction and analysis by obtaining a multi-modal terrain map of a target star catalog area, including multi-dimensional information such as the geometric characteristics and mechanical characteristics of the terrain. At the same time, the current state and candidate motion primitives of the star vehicle are combined, the dynamic information of the vehicle itself is associated with the terrain information, the interaction between the star vehicle and the terrain under different states and motion modes is fully considered, and the information related to the terrain and vehicle motion is processed and transformed respectively. The shared feature reflects the comprehensive characteristics of the terrain, and the combined encoding result reflects the motion characteristics of the vehicle under a specific motion primitive. The shared feature and the combined encoding result are mutually related and complementary, providing effective and representative feature representations for subsequent feature fusion. The combined encoding result and the shared feature are then dimensionally matched and spliced to obtain a one-dimensional fusion feature tensor, realizing the deep fusion of terrain information and vehicle motion information at the feature level. The fused feature tensor contains comprehensive information of the terrain and vehicle motion, and finally the loss predictor predicts the loss according to the fused feature tensor to obtain the movement cost of the star vehicle executing the candidate motion primitive.

[0016] By integrating multi-modal terrain maps, the current state of the star vehicle, and candidate motion primitives, etc. Multi-source information can more comprehensively represent the interaction between the star vehicle and the complex star environment. The shared feature of the terrain geometry and mechanics enables the model to not only focus on the surface geometry of the terrain, but also consider the influence of the mechanical properties of the terrain on the movement of the star vehicle, avoiding false judgments of geometrically feasible but mechanically infeasible situations, thereby improving the accuracy of the passability prediction of the star vehicle. Moreover, the present application uses a passability prediction model for unified processing, organically combining feature extraction, motion encoding, feature fusion, and loss prediction processes, avoiding the dependence on soil mechanical parameters that are difficult to obtain in real time in traditional vehicle dynamics-based methods, reducing computational complexity, and speeding up the reasoning speed. Real-time prediction can be achieved, providing a more timely and reliable basis for the safe and efficient movement of the star vehicle.

[0017] In summary, the present application introduces shared features of terrain geometry and mechanics to make up for the shortcomings of relying solely on geometric parameters for prediction, enabling the identification of dangerous terrain that is geometrically feasible but mechanically infeasible, thereby improving the reliability of the prediction. The passability prediction model used in the present application avoids direct dependence on these complex parameters, simplifies the calculation process by learning and fusing multi-modal information and vehicle state and motion information, greatly improving the prediction efficiency, and providing a strong guarantee for the safe and efficient movement of the star vehicle in complex star environments. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The flowchart of the passability prediction method of the star vehicle of the present application; Figure 2 Fig. 1 is a schematic diagram of a kinematic model of a six-wheel star car according to an embodiment of the present application; Figure 3 Fig. 2 is a schematic diagram of a sampling simulation process of a star car according to an embodiment of the present application; Figure 4 Fig. 3 is a schematic diagram of a model network structure according to an embodiment of the present application. DETAILED DESCRIPTION

[0019] In order to make the above objectives, features and advantages of the present application more apparent, specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present application are shown in the drawings, it is understood that the present application can be implemented in various forms, and should not be interpreted as being limited to the embodiments set forth herein, but rather, these embodiments are provided so as to more thoroughly and completely understand the present application. It is understood that the drawings and embodiments of the present application are for illustrative purposes only, and are not intended to limit the scope of the present application.

[0020] It is understood that each of the steps recited in the method embodiments of the present application can be performed in different orders, and / or in parallel. In addition, the method embodiments can include additional steps and / or omit performing the steps shown. The scope of the present application is not limited in this respect.

[0021] The term "comprising" and variations thereof as used herein are used inclusively, i.e., "comprising, but not limited to"; the term "based on" is "based at least in part on"; the term "one embodiment" is "at least one embodiment"; the term "another embodiment" is "at least one additional embodiment"; the term "some embodiments" is "at least some embodiments"; the term "optionally" is "optional embodiments." Related definitions of other terms will be given in the description below. It is noted that the concepts of "first", "second", etc. mentioned in the present application are merely used to distinguish different devices, modules or units, and are not intended to limit the order or interdependence of the functions performed by these devices, modules or units.

[0022] It is noted that the modification of "one" or "more" mentioned in the present application is illustrative rather than limiting, and those skilled in the art should understand that, unless otherwise explicitly stated in the context, it should be understood as "one or more".

[0023] The names of the messages or information exchanged between the plurality of devices in the embodiments of the present application are merely used for illustrative purposes, and are not intended to limit the scope of the messages or information.

[0024] It should be noted that the information (including but not limited to user equipment information, user personal information, etc.), data (including but not limited to data for analysis, stored data, displayed data, etc.) and signals involved in the present application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data need to comply with relevant laws, regulations and standards of relevant countries and regions, and provide corresponding operation portal for user to choose authorization or rejection.

[0025] In combination Figure 1 The star vehicle passability prediction method provided by the embodiment of the present application comprises: Obtaining a multi-modal terrain map of a target star catalog region, a current state of a star vehicle and candidate motion primitives.

[0026] Specifically, the multi-modal terrain map of the target star catalog region includes various terrain data such as an elevation map, a ground friction characteristic map and a ground softness characteristic map. These data respectively represent the geometric information, shear characteristics and bearing characteristics of the terrain. Specifically, the elevation map is generated by a digital elevation model (DEM) or point cloud data, and can provide geometric parameters such as the slope and undulation of the terrain; the ground friction characteristic map reflects the shear strength of the ground surface based on the normalized internal friction angle; and the ground softness characteristic map reflects the bearing performance of the ground surface by using the normalized equivalent stiffness modulus. The current state of the star vehicle includes recording key parameters such as its position and heading angle, and in some preferred embodiments, it is usually realized by the positioning and navigation sensors carried on the vehicle. The position information is determined by a star catalog positioning system, and the heading angle is measured according to an inertial measurement unit (IMU). The current state provides the instantaneous position and direction information of the star vehicle for the prediction model, helping the model to determine which motion primitive is feasible under a specific terrain. The candidate motion primitives are determined based on the kinematic model and control strategy of the star vehicle. The motion primitives define a series of basic motion actions that the star vehicle can take, such as forward movement, backward movement, turning, etc.

[0027] These primitives are usually represented in the form of parameters, for example, a turning angle δ and a moving direction , δ a random sampling between the maximum turning angle to , 1 or -1 representing forward movement or backward movement. By discretizing the control quantity, a series of motion primitives can be generated to provide a basis for subsequent path planning.

[0028] The multi-modal terrain map, the current state and the candidate motion primitives are input into a passability prediction model.

[0029] Specifically, in a preferred embodiment of the present application, the prediction model can be deployed in the control system of the planetary vehicle, receiving the above-mentioned input information through a data interface, and the multi-modal terrain map is input in a specific data format (such as a multi-channel tensor), and the current state and the candidate motion primitive are converted into corresponding numerical vector forms and input into the model.

[0030] The multi-modal terrain map is subjected to feature extraction by a feature extractor of the passability prediction model, to obtain shared features of terrain geometry and mechanics coupling.

[0031] Specifically, the feature extractor is constructed based on a convolutional neural network, including multiple convolutional layers and pooling layers, and the elevation map, the ground friction characteristic map and the ground softness characteristic map are subjected to convolutional operation to extract preliminary features, and then subjected to feature interaction and fusion between channels to form a shared feature map that comprehensively reflects the terrain geometry and mechanical characteristics. For example, a convolution kernel slides on the elevation map to calculate the height change features of a local area, while the friction coefficient change and softness degree change features of the corresponding areas are extracted from the friction characteristic map and the softness characteristic map, and then these features are spliced and fused to obtain shared features that can reflect the coupling relationship between terrain geometry and mechanics.

[0032] The current state of the planetary vehicle and the candidate motion primitive are combined and encoded by a motion encoder of the passability prediction model, to obtain a combined encoding result.

[0033] Specifically, the motion encoder adopts a multi-layer neural network structure to combine the current state (position coordinates and heading angle) and the candidate motion primitive (turning angle and moving direction) into an input vector. After transformation by a fully connected layer and an activation function, the input vector is mapped to a fixed-length encoding vector, which is the combined encoding result and can represent the potential motion features of different motion primitives under the current state.

[0034] The combined encoding result and the shared features are subjected to dimension matching and splicing to obtain a one-dimensional fusion feature tensor.

[0035] Specifically, when the combined encoding result and the shared features are subjected to dimension matching and splicing, the dimensions of the two are adjusted so that they can be spliced in the same dimension. Through flattening and concatenation operations, the output of the motion encoder and the output of the feature extractor are merged into a one-dimensional fusion feature tensor, forming a feature vector that comprehensively integrates terrain and motion information.

[0036] The one-dimensional fusion feature tensor is subjected to loss prediction by a loss predictor of the passability prediction model, to obtain the movement cost of the planetary vehicle when executing the candidate motion primitive.

[0037] Specifically, loss prediction is performed using a loss predictor. The fused feature tensor is input into the loss predictor, which contains multiple branches that predict movement costs such as time consumption, energy consumption, and danger level. The prediction results represent the expected cost of each candidate motion primitive on a specific terrain, providing a quantitative basis for path planning decisions.

[0038] The rover accessibility prediction method in this embodiment acquires multimodal terrain maps of the target star surface region, including multi-dimensional information such as the geometric and mechanical features of the terrain, providing richer basic data for subsequent feature extraction and analysis. Simultaneously, by combining the rover's current state and candidate motion primitives, the method correlates the vehicle's dynamic information with terrain information, fully considering the interaction between the rover and the terrain under different states and motion modes. Furthermore, it performs targeted processing and transformation on terrain and vehicle motion-related information. Shared features reflect the comprehensive characteristics of the terrain, while the combined encoding results reflect the vehicle's motion characteristics under specific motion primitives. The shared features and combined encoding results are interconnected and complementary, providing an effective and representative feature representation for subsequent feature fusion. The combined encoding results and shared features are then dimensionally matched and concatenated to obtain a one-dimensional fused feature tensor, achieving deep fusion of terrain and vehicle motion information at the feature level. The fused feature tensor contains comprehensive information on terrain and vehicle motion. Finally, a loss predictor performs loss prediction based on the fused feature tensor to obtain the movement cost of the rover executing candidate motion primitives.

[0039] By integrating multi-modal terrain maps, the rover's current state, and candidate motion primitives, the interaction between the rover and the complex planetary environment can be more comprehensively characterized. The shared characteristics of terrain geometry and mechanics coupling allow the model to consider not only the surface geometry of the terrain but also the impact of its mechanical properties on rover movement, avoiding misjudgments of geometrically feasible but mechanically infeasible scenarios, thus improving the accuracy of rover accessibility prediction. Furthermore, this embodiment employs a unified accessibility prediction model, organically combining feature extraction, motion encoding, feature fusion, and loss prediction processes. This avoids the reliance on soil mechanics parameters, which are difficult to obtain in real time, as in traditional vehicle dynamics-based methods, reducing computational complexity, accelerating inference speed, and enabling real-time prediction. This provides a more timely and reliable basis for the safe and efficient movement of the rover.

[0040] In summary, this embodiment overcomes the limitations of predictions relying solely on geometric parameters by introducing shared features that couple terrain geometry and mechanics. It can identify geometrically feasible but mechanically infeasible hazardous terrain, thereby improving prediction reliability. The accessibility prediction model employed in this embodiment avoids direct dependence on these complex parameters. By learning and fusing multimodal information, as well as vehicle state and motion information, it simplifies the calculation process, significantly improves prediction efficiency, and provides strong support for the safe and efficient movement of planetary rovers in complex asteroid environments.

[0041] Optionally, acquiring the multimodal terrain map of the target star surface region, the current state of the rover, and candidate motion primitives includes: Using the current position of the rover as the center, obtain a local elevation map, a ground friction characteristic map, and a ground softness characteristic map within a preset range of the rover; The local elevation map, the ground friction characteristic map, and the ground softness characteristic map are stacked by channels to generate the multimodal terrain map; Obtain the x-coordinate and y-coordinate of the center of the planetary rover in a preset coordinate system, as well as the heading angle of the longitudinal axis of the planetary rover relative to the y-axis of the preset coordinate system. The current state is formed based on the horizontal coordinate, the vertical coordinate, and the heading angle; The candidate motion primitives are generated based on the preset steering angle and preset direction of movement of the planetary rover.

[0042] Specifically, in a preferred embodiment of the present invention, the local elevation map can be acquired by a lidar or high-resolution camera mounted on the rover. These devices can scan the planetary surface and measure the elevation information within a preset range centered on the current location, forming a two-dimensional elevation data matrix. The ground friction characteristic map and the ground softness characteristic map can be generated based on the physical property data of the planetary surface soil. In an optional embodiment of the present invention, this data can be obtained through a pre-stored planetary surface soil model or through real-time measurement by onboard sensors, or by obtaining corresponding semantics through camera image segmentation, and then mapping the semantics to physical properties, again centered on the current location and covering the same preset range, forming the friction characteristic matrix and the softness characteristic matrix respectively. The elevation data matrix, friction characteristic matrix, and softness characteristic matrix are each treated as three independent channels and stacked together in a certain order to form a three-dimensional data tensor, i.e., a multimodal topographic map. This stacking method is similar to a multi-channel image in image processing, where each channel represents a specific type of terrain information, so that the subsequent feature extractor can process these different types of terrain data simultaneously. The center position of the rover can be determined by its own positioning system, obtaining its x and y coordinates in a preset coordinate system (usually a reference coordinate system on the planet's surface). The heading angle is measured by the rover's inertial measurement unit (IMU) or attitude sensor, representing the angle between the rover's longitudinal axis and the y-axis of the preset coordinate system, thus characterizing the rover's orientation. The current state is composed of the x, y, and heading angles. Combining these three parameters into a vector constitutes the rover's current state vector, which concisely describes the rover's position and orientation in the preset coordinate system, providing basic state information for subsequent motion planning and passability prediction. The preset steering angle is a set of possible steering angles pre-set according to the rover's design parameters and mobility capabilities, for example, setting a steering gear at regular intervals. The preset movement direction includes two basic directions: forward and backward. By combining each preset steering angle with each preset movement direction, a series of candidate motion primitives are generated. Each motion primitive represents a possible movement mode that the rover can attempt in its current position and state, used for subsequent motion simulation and cost evaluation.

[0043] In a preferred embodiment of the present invention, the multimodal terrain map M is a robot-centric multichannel tensor, including a local elevation map H, a two-dimensional matrix representing terrain geometry; and a ground friction characteristic map F, a two-dimensional matrix representing surface shear characteristics, which can be represented by a normalized internal friction angle. φ The diagram S represents the soft surface characteristics of the ground, a two-dimensional matrix that characterizes the bearing capacity of the surface and can be represented by the normalized equivalent stiffness modulus N. The robot's current posture is also shown. s i Characterized as s i=( ) , representing the position coordinates and heading angle, respectively. Motion primitives m i Characterized as m i = T ,in, δ i For steering angle, d i Indicates the direction of movement.

[0044] In this optional embodiment, by acquiring a local elevation map, ground friction characteristic map, and ground softness characteristic map within a preset range centered on the rover's current position, the terrain features of the area to be traversed can be accurately depicted, providing a detailed data foundation for subsequent analysis. Simultaneously, stacking the three different types of terrain maps by channel generates a multimodal terrain map, enabling the model to process multiple terrain information simultaneously, fully considering the geometric and mechanical properties of the terrain, and overcoming the problem of distorted prediction results caused by relying solely on a single terrain feature. Acquiring the rover's x-coordinate, y-coordinate, and heading angle in a preset coordinate system and forming its current state provides the model with precise position and orientation information of the rover in space, facilitating better planning of subsequent movements. Generating candidate motion primitives based on preset steering angles and directions of movement provides the rover with multiple possible motion options, ensuring the diversity and feasibility of motion planning and aiding in finding the optimal motion path.

[0045] Optionally, the feature extractor of the accessibility prediction model extracts features from the multimodal terrain map to obtain shared features coupling terrain geometry and mechanics, including: The local elevation map, the ground friction characteristic map, and the ground softness characteristic map from the multimodal topographic map are respectively input into the multi-channel of the feature extractor; Through the convolutional structure of the feature extraction layer, the multi-channel input terrain data is subjected to layer-by-layer feature extraction to obtain the geometric features in the local elevation map, the mechanical property distribution features in the ground friction characteristic map and the ground softness characteristic map; The feature fusion module of the feature extraction layer couples the geometric features with the mechanical property distribution features across channels to generate the shared features.

[0046] Specifically, in combination Figure 4As shown, the accessibility prediction model includes a feature extractor. The feature extractor is designed as a convolutional structure, which calculates the shared features of the entire map M for different motion primitives, demonstrating a certain computational efficiency advantage. The local elevation map, ground friction characteristic map, and ground softness characteristic map from the multimodal terrain map are respectively input into the multi-channel feature extractor. Each channel of the multimodal terrain map corresponds to a specific type of terrain information; that is, the local elevation map, ground friction characteristic map, and ground softness characteristic map are three independent input channels. This ensures that each type of terrain information is processed independently and completely during feature extraction, preserving its own characteristics. The convolutional structure in each channel consists of a series of convolutional layers and pooling layers. For the local elevation map channel, the convolutional layers use convolutional kernels of different sizes and types to extract the geometric features of the terrain layer by layer, such as edges and slope changes. Similarly, for the ground friction characteristic map and ground softness characteristic map channels, the convolutional layers extract the corresponding mechanical property distribution features, such as the changing trend of the friction coefficient and the spatial distribution of softness. Pooling layers are used to reduce data dimensionality, highlight important features, and improve computational efficiency. The feature fusion module can employ various methods, such as stitching together the feature maps extracted from each channel and then further fusing them through convolutional layers, or using attention mechanisms to emphasize the important correlations between different features. The resulting shared features encompass both the geometric information of the terrain and its mechanical properties, providing a more comprehensive and accurate terrain description for subsequent motion planning and cost assessment.

[0047] In this optional embodiment, the multimodal topographic map employs a multi-channel input method, ensuring that local elevation maps, ground friction characteristic maps, and ground softness characteristic maps can be processed independently and specifically. This allows the feature extractor to accurately extract unique key information from each type of terrain data, laying a solid foundation for subsequent feature fusion. Secondly, the layer-by-layer feature extraction using a convolutional structure can gradually extract more abstract and advanced features from the original terrain data, making the extraction of geometric and mechanical property distribution features more accurate and hierarchical. It gradually integrates simple geometric features such as local slopes and edges into a comprehensive understanding of the entire terrain region, while accurately capturing mechanical properties such as the changing trend of friction coefficients and the spatial distribution of softness. Finally, the feature fusion module, through cross-channel coupling, organically combines geometric features and mechanical property distribution features, breaking the isolation between different modal information. This allows the geometry and mechanical properties of the terrain to complement and corroborate each other, generating shared features that comprehensively reflect the geometry and mechanical properties of the terrain. This method effectively solves the problem of distorted prediction results caused by traditional methods relying on only a single terrain feature. It ensures that the prediction results can more realistically reflect the impact of complex terrain on the rover's passage, providing high-quality, comprehensive and reliable feature input for subsequent movement cost prediction. This significantly enhances the rover's ability to predict passability in unknown and complex star surface environments.

[0048] Optionally, the motion encoder using the accessibility prediction model combines the current state of the rover with the candidate motion primitives to obtain a combined encoding result, including: The horizontal coordinate, vertical coordinate, and heading angle in the current state are combined with the preset turning angle and preset movement direction in the candidate motion primitive to form a motion-state joint parameter set; The motion-state joint parameter set is numerically mapped and dimensionally transformed by the fully connected layer of the motion encoder to obtain the feature vector of the motion-state joint parameter set. The feature vector is normalized to obtain the combined encoding result.

[0049] Specifically, the rover's x-coordinate and y-coordinate in a preset coordinate system are obtained from its positioning and navigation system, and the heading angle of the vehicle's longitudinal axis relative to the y-axis of the preset coordinate system is obtained through the inertial measurement unit (IMU). Simultaneously, preset sets of steering angles and movement directions are acquired. Then, these parameters—x-coordinate, y-coordinate, heading angle, preset steering angle, and preset movement direction—are combined into a joint parameter set. This joint parameter set comprehensively describes the rover's integrated information under its current state and possible movement intentions, providing complete input data for subsequent encoding processing. The fully connected layer in the motion encoder consists of multiple neurons, each connected to various parameters in the input parameter set. When the motion-state joint parameter set is input to the fully connected layer, the neurons perform a weighted summation to numerically map the input parameters. The weights are learned by the encoder during training. This nonlinear transformation converts the original parameters into a fixed-length feature vector. This feature vector resides in a new feature space, which can more efficiently represent the intrinsic characteristics of the motion-state joint information. Simultaneously, the dimensionality transformation also facilitates unified processing by subsequent modules. Normalization involves scaling each element of the feature vector to standardize its numerical range. Common methods include min-max normalization and Z-score normalization. After normalization, the elements of the feature vector are on the same scale, eliminating the influence of differences in parameter units and numerical ranges. This allows the feature vector to more accurately reflect the relative relationships and inherent structure of the motion-state joint information. Ultimately, this normalized feature vector is the combined encoding result, fusing the current state of the rover and the information of candidate motion primitives into a compact and robust representation, providing high-quality input for subsequent fusion processing and loss prediction. Figure 4As shown, the accessibility prediction model includes a motion encoder. The robot state and motion primitives are first encoded by combination, and then converted into tensors by the motion encoder. The feature encoder encodes the height and physical information layers, where the physical information layer is obtained by merging the stiffness and friction layers.

[0050] In this optional embodiment, by combining the horizontal coordinate, vertical coordinate, heading angle, and preset steering angle and direction of movement into a joint motion-state parameter set, a comprehensive integration of the rover's current state and potential motion intentions is achieved. This combination ensures that subsequent processing is based on a comprehensive information set, avoiding information fragmentation and omission, and providing a more complete and accurate input for subsequent feature extraction and encoding. Furthermore, the fully connected layer of the motion encoder performs numerical mapping and dimensional transformation on the joint motion-state parameter set, converting the original parameters into fixed-length feature vectors. This not only reduces data complexity but also highlights key features in the data, enabling the feature vectors to more efficiently represent the inherent characteristics of the joint motion-state information. Simultaneously, dimensional transformation facilitates unified processing by subsequent modules, improving computational efficiency and model compatibility. Finally, the feature vectors are normalized, further optimizing their representation. Normalization ensures that all elements of the feature vectors are on the same scale, eliminating the influence of differences in parameter dimensions and numerical ranges, and enhancing the stability and comparability of the feature vectors. This makes the combined encoding results more accurately reflect the relative relationships and inherent structure of motion-state joint information, providing high-quality input for subsequent fusion processing and loss prediction.

[0051] Optionally, the step of performing dimensional matching and concatenation of the combined encoding result and the shared features to obtain a one-dimensional fused feature tensor includes: The shared features are flattened to convert the feature matrix of the shared features into a one-dimensional feature vector; The combined encoding result and the one-dimensional feature vector are concatenated in a first preset order to obtain the one-dimensional fused feature tensor.

[0052] Specifically, shared features typically exist in the form of a multidimensional matrix, containing coupled information of terrain geometry and mechanics. A flattening operation is performed on the shared features to effectively concatenate and fuse them with the combined encoding results. This flattening operation involves rearranging the multidimensional feature matrix in a certain order (such as row-major or column-major) into a one-dimensional feature vector.

[0053] For example, suppose the shared feature is a m × n A two-dimensional matrix can be transformed into a matrix of length by flattening. m × nThe one-dimensional vector is used so that each element has a corresponding index position in the vector, which preserves all the information in the original feature matrix and transforms the data organization to prepare for the subsequent concatenation operation.

[0054] The combined encoding result is then concatenated with the one-dimensional feature vector in a first preset order to obtain a one-dimensional fused feature tensor. The combined encoding result is a normalized feature vector containing fused information about the rover's current state and candidate motion primitives. This combined encoding result feature vector is then concatenated with the flattened shared feature vector in a preset order, such as placing the combined encoding result first, then the shared feature vector, or vice versa. The concatenation operation links the two vectors end-to-end to form a new, longer one-dimensional vector, namely the one-dimensional fused feature tensor.

[0055] For example, if the combined encoding result is a length of p The flattened shared feature of the vector is a vector of length . q The length of the concatenated fused feature tensor is _ ... p+q This fusion feature tensor integrates information from terrain features and motion state, providing comprehensive input to the subsequent loss predictor. This allows loss prediction to simultaneously consider terrain conditions and the rover's motion state, thereby achieving a more accurate assessment of movement costs.

[0056] In this optional embodiment, the shared features are flattened into a one-dimensional feature vector, preserving terrain information while changing the data organization to allow for concatenation with the combined encoding results. Secondly, the combined encoding results and the flattened shared features are concatenated in a preset order to generate a one-dimensional fused feature tensor. This tensor integrates the rover's state and terrain information, providing comprehensive input for subsequent loss prediction. This ensures that the prediction simultaneously considers terrain conditions and the rover's motion state, improving the accuracy of movement cost assessment and enhancing the rover's adaptability to complex planetary environments.

[0057] Optionally, the loss predictor of the accessibility prediction model, based on the one-dimensional fusion feature tensor, performs loss prediction to obtain the movement cost of the rover when executing the candidate motion primitive, including: The one-dimensional fusion feature tensor is input into the shared fully connected layer of the loss predictor to extract high-dimensional fusion features; The high-dimensional fusion features are respectively input into multiple parallel branches of the loss predictor to perform loss prediction, and the predicted value corresponding to each parallel branch is obtained. The movement cost is generated based on the predicted value of each of the parallel branches.

[0058] Specifically, the shared fully connected layer consists of multiple neurons, each connected to an element of the input feature tensor. When the one-dimensional fused feature tensor is input, the shared fully connected layer transforms the input features through weighted summation and nonlinear activation functions (such as ReLU) to extract higher-level fused features. This process captures the complex relationships between different elements in the one-dimensional fused feature tensor, further enhancing the expressive power of the features and making loss prediction more accurate. Simultaneously, the loss predictor in this embodiment is designed with multiple parallel branches, each specifically responsible for predicting a particular loss type. For example, one branch predicts time cost, another predicts energy cost, and a third predicts danger cost. Each branch contains several layers of neural network structures, such as fully connected layers and activation functions, used to extract features related to the corresponding loss type from the high-dimensional fused features and output a preliminary prediction value for that loss type. After obtaining the preliminary prediction values ​​from each parallel branch, these prediction values ​​are integrated into the final movement cost through a specific fusion strategy (such as weighted summation or direct concatenation). The movement cost is a comprehensive metric used to evaluate the overall cost when the rover executes candidate motion primitives. For example, different weights can be assigned to time, energy, and risk levels, adjusting the influence of each cost type in the final movement cost according to task requirements. The resulting movement cost comprehensively reflects the overall cost of executing the movement primitive, providing crucial information for subsequent path planning and decision-making.

[0059] In a preferred embodiment of the invention, the movement cost c Characterized as c = T This is considered from three aspects: time, energy, and degree of danger. Among them, The time spent executing the motion primitive; The energy consumed is obtained by summing the integrals of the torque of each drive motor over time; The degree of danger is used to reflect the probability or level of risk of slipping, sinking, etc. when performing the sport.

[0060] In summary, combining Figure 4 As shown, the passability prediction model includes a loss predictor. The output tensors of both the motion encoder and the feature encoder are flattened into a 1-dimensional implicit encoding structure, merged and concatenated in terms of dimension to obtain a one-dimensional feature tensor, which serves as the input to the loss predictor. The loss predictor first uses convolution, then splits into three branches: time, energy, and risk level, predicting the corresponding passability loss for each.

[0061] In this optional embodiment, by inputting the one-dimensional fusion feature tensor into a shared fully connected layer to extract high-dimensional fusion features, deeper information within the fusion features can be further mined, enhancing the expressive power of the features. The weighted summation and nonlinear activation operations of the shared fully connected layer can capture the complex relationships between different feature elements, enriching the feature representation and providing more discriminative input for subsequent loss prediction. Secondly, using multiple parallel branches to predict the loss of the high-dimensional fusion features enables accurate evaluation of different types of losses simultaneously. Each branch performs specialized feature extraction and prediction for a specific loss type (such as time, energy, or risk level). This division of labor not only improves the targeting and accuracy of the prediction but also allows for parallel processing of multiple loss types, improving model efficiency. Finally, by synthesizing the prediction values ​​from each parallel branch to generate the movement cost, a comprehensive and integrated evaluation result can be provided. This approach integrates losses from different dimensions such as time, energy, and risk level into an indicator that reflects the overall cost. Furthermore, in practical applications, the weights of the prediction values ​​from each branch can be flexibly adjusted according to task requirements, making the movement cost more aligned with specific task objectives and priorities.

[0062] Optionally, the parallel branches include a time branch, an energy branch, and a danger level branch. The step of inputting the high-dimensional fused features into multiple parallel branches of the loss predictor for loss prediction, to obtain the predicted value corresponding to each parallel branch, includes: The high-dimensional fusion features are mapped through the first fully connected layer of the time branch to obtain the predicted time cost required to execute the candidate motion primitive, and the predicted time cost is used as the predicted value of the time branch. The high-dimensional fusion features are mapped through the second sub-fully connected layer of the energy branch to obtain the predicted energy cost required to execute the candidate motion primitive, and the predicted energy cost is used as the predicted value of the energy branch. The high-dimensional fusion features are mapped through the third fully connected layer of the danger level branch to obtain the predicted danger level value when executing the candidate motion primitive, and the predicted danger level value is used as the predicted value of the danger level branch.

[0063] Specifically, the first sub-fully connected layer consists of multiple neurons, each connected to elements of the high-dimensional fused feature. The features are transformed using weighted summation and non-linear activation functions (such as ReLU) to extract features related to time cost. For example, the first sub-fully connected layer learns the association between certain terrain features (such as steep slopes or soft areas) and time cost, thus predicting the time required to execute the motion primitive. Similarly, the second sub-fully connected layer processes the high-dimensional fused feature using weighted summation and non-linear activation functions to extract features related to energy consumption. For example, the energy branch learns that energy consumption is higher in high-friction areas or on paths requiring frequent turning, thus predicting the energy cost under different motion primitives. The third sub-fully connected layer also employs weighted summation and non-linear activation functions, focusing on extracting features related to hazard level. For example, the hazard level branch identifies the association between certain terrain features (such as steep slopes, soft edges, or areas with dense obstacles) and high hazard level, thus predicting the level of risk the vehicle may face when executing a specific motion primitive.

[0064] In this optional embodiment, the time branch maps high-dimensional fusion features through the first fully connected layer to accurately predict the predicted time cost required to execute candidate motion primitives. This branch focuses on extracting time-dimensional features, taking into account factors such as terrain complexity, vehicle status, and the rationality of motion commands to provide a time reference for task planning. The energy branch maps through the second fully connected layer to obtain the predicted energy cost. This branch focuses on energy consumption-related features, providing key data for resource management and path optimization based on terrain resistance, motion intensity, and vehicle energy consumption characteristics. The danger level branch maps through the third fully connected layer to obtain the predicted danger level value. This branch focuses on potential risk features, combining terrain treacherousness, vehicle stability, and passability to quantify risks in advance and ensure driving safety. All branches work together to provide comprehensive and accurate cost information for subsequent decisions.

[0065] Optionally, generating the movement cost by concatenating or weighting the predicted values ​​of each of the parallel branches includes: The movement cost is obtained by vector combination of the predicted time cost, predicted energy cost, and predicted danger level value corresponding to the time branch, the energy branch, and the danger level branch, respectively, in a second preset order. Alternatively, corresponding weights can be assigned to the prediction time cost, the prediction energy cost, and the prediction risk level value, respectively; The movement cost is obtained by weighting and fusing the predicted time cost, the predicted energy cost, and the predicted danger level value according to their respective weights.

[0066] In a preferred embodiment of the present invention, two optional methods are provided when generating the movement cost: vector concatenation and weighted fusion. Both methods operate based on the predicted values ​​corresponding to the time branch, energy branch, and danger level branch, respectively.

[0067] Specifically, the first method is vector concatenation. Following a pre-defined order, the predicted time cost from the time branch, the predicted energy cost from the energy branch, and the predicted danger level from the danger level branch are combined into a vector. This vector represents the movement cost, simultaneously reflecting the impact of the three key factors—time, energy, and danger level—on movement. In this way, the movement cost is presented in a structured manner, allowing subsequent path planning and decision-making processes to comprehensively consider information from multiple dimensions.

[0068] The second method is weighted fusion. First, a weight is assigned to each of the predicted time cost, predicted energy cost, and predicted hazard level. These weights can be set according to the actual task requirements and priorities to reflect the importance of different factors in the task. Then, based on these weights, the predicted time cost, predicted energy cost, and predicted hazard level are weighted and summed to obtain a comprehensive movement cost. This method allows for trade-offs among different factors, making the final movement cost more closely aligned with the specific goals and constraints of the task, providing a single metric for path planning that comprehensively considers time, energy, and hazard level.

[0069] In this embodiment of the invention, by concatenating or weighting the predicted values ​​corresponding to the time branch, energy branch, and danger level branch, the generated movement cost can comprehensively reflect information from multiple key dimensions. This multi-dimensional cost representation provides a more comprehensive basis for path planning, enabling path planning to consider not only time efficiency but also energy consumption and safety.

[0070] For example, in some tasks, although a path may be shorter in time, it may have higher energy consumption or a higher degree of danger. By comprehensively evaluating the costs from multiple dimensions, such paths can be avoided, thereby improving the overall efficiency and safety of the task.

[0071] Furthermore, the embodiments of the present invention provide two methods: vector concatenation and weighted fusion, enabling the present invention to flexibly adapt to different task requirements and scenarios.

[0072] For example, for some tasks, it is necessary to retain independent information for each dimension in order to conduct detailed analysis and decision-making, in which case vector concatenation schemes are more suitable. For other tasks, a comprehensive cost metric is needed for rapid comparison and optimization, in which case weighted fusion schemes are more appropriate.

[0073] By employing a weighted fusion approach, weights can be assigned to the costs of different dimensions based on the specific requirements of the task. For example, in resource-constrained tasks, the weight of energy cost can be increased to prioritize paths with lower energy consumption; in time-critical tasks, the weight of time cost can be increased to select the shortest path. This weighting mechanism enables path planning to more accurately meet the optimization objectives of the task, improving its efficiency and effectiveness. Furthermore, the vector concatenation method preserves the independent information of each dimension, allowing the model to more comprehensively evaluate path feasibility in complex environments. Even if the predicted value of one dimension is affected by noise or error, information from other dimensions can still provide effective supplementation, thereby improving the model's robustness in complex environments. In addition, the weighted fusion method, by integrating information from multiple dimensions, can also reduce the impact of single-dimensional errors on the final decision, further improving the model's robustness. The vector concatenation scheme, by explicitly representing the predicted value of each dimension, makes the model's output more interpretable. Path planners can directly view the cost of each dimension, understand the specific basis for path selection, and thus make better decisions and adjustments. This interpretability is particularly important for decision support in complex tasks, helping to improve the transparency and credibility of the task.

[0074] In this optional embodiment, the move cost is generated by vector concatenation or weighted fusion based on the predictions of each parallel branch, which not only provides multi-dimensional decision support but also enhances the model's flexibility, adaptability, robustness, and interpretability.

[0075] Optionally, the planetary vehicle accessibility prediction method further includes: constructing the accessibility prediction model; The construction of the accessibility prediction model includes: In a simulated star table environment, multimodal terrain maps, current states, candidate motion primitives, and the actual time cost, energy cost, and hazard labels corresponding to the candidate motion primitives are collected under different terrain conditions to generate a training sample set. The training sample set is subjected to topographic map random rotation processing, height drift processing, and noise perturbation processing to obtain an enhanced training sample set; The enhanced training sample set is input into the initial convolutional neural network, and the initial convolutional neural network is trained using a multi-task loss function; The initial convolutional neural network after training is used as the accessibility prediction model.

[0076] Specifically, firstly, a large amount of data was collected in a simulated star table environment. This data included multimodal terrain maps of the rover under different terrain conditions. These maps contained not only geometric information (such as slope and undulation) but also mechanical properties (such as friction coefficient and equivalent stiffness). Simultaneously, the rover's current state, such as its position and heading angle, as well as its potential candidate motion primitives, such as steering angle and direction of movement, were recorded. Furthermore, for each candidate motion primitive, its corresponding actual time cost, energy cost, and hazard label were recorded. These labels reflected the risks faced when executing the motion primitive, such as slippage or sinking. In this way, a training sample set rich in information was generated, providing a solid data foundation for model training. Further, the terrain map was randomly rotated to simulate the terrain changes the rover might encounter when moving in different directions, enabling the model to better adapt to terrain features in different directions. Secondly, height drift processing was performed. By adding random offsets to the height information of the terrain map, the model became less sensitive to the absolute height of the terrain, thus focusing more on the relative changes in terrain. Finally, noise perturbation was added to the terrain map to increase data diversity, making the model more robust to measurement or map reconstruction errors that may occur in real-world applications. These processes resulted in an enhanced training sample set, providing higher-quality data for subsequent model training. An initial convolutional neural network (CNN) was trained using this enhanced training sample set. This CNN can handle multiple tasks simultaneously, namely predicting time cost, energy cost, and danger level. The network was trained using a multi-task loss function that comprehensively considers prediction errors for time, energy, and danger level. This method in this embodiment allows the network to learn a unified feature representation while optimizing multiple related prediction tasks. During training, the network gradually learns how to extract useful information from the input multimodal terrain map and rover state and map it to the corresponding cost predictions. In this way, the trained CNN can accurately predict the cost of new, unprocessed combinations of terrain and state. After the above training process, the initial CNN has the ability to accurately predict the movement cost of the rover on different terrains. Therefore, this trained network is used as the final accessibility prediction model. This model can take new multimodal terrain maps and rover status as input and output corresponding predicted values ​​for time cost, energy cost, and hazard level. These predictions provide important basis for the rover's path planning and decision-making, helping it select the optimal movement path in the complex planetary environment, thereby improving the success rate and efficiency of the mission.

[0077] In a preferred embodiment of the present invention, combined with Figure 4 As shown, the detailed process of the planetary rover performing each sampling segment in the simulation environment includes: Initial stage: In At time 0, the rover is initialized to a random position in the simulation environment. Initialization includes setting the rover's initial pose and heading angle. Stationary Phase: After initialization, the rover enters a stationary phase for a duration Δ. During this period, the rover remains stationary to stabilize its initial position on the ground. The purpose of this stationary phase is to ensure that the rover has adapted to the terrain before starting to move and to accurately record its initial state. Registration initialization information: After the stationary phase ends, the system records the rover's initial state and relevant terrain information. This information includes the terrain height map H and physical property map P, as well as the rover's current attitude. s Movement segment: The planetary rover moves according to preset motion primitives. Move, duration Δ Movement primitives This includes steering angle and direction of movement. During the movement phase, the rover performs specific motions, while the system records its state changes and terrain interaction information. Termination phase: The simulation... The simulation may terminate for various reasons, including the rover successfully completing its movement, a collision occurring, excessive wheel slippage, excessive wheel sinking, exceeding the simulation time limit, or reaching the end of the path. Upon simulation termination, the system calculates and records the losses incurred by the rover during that segment of movement. This includes time cost, energy cost, and level of danger. Losses are calculated and data is stored: based on the rover's performance during the movement phase, losses are calculated. This includes time cost, energy cost, and risk level. The calculated loss and related data are saved for subsequent model training and analysis. This embodiment can generate a large number of training samples in the simulation environment. These samples include the performance of the planetary rover under different terrain and motion conditions, providing rich data support for building an accurate accessibility prediction model.

[0078] In a preferred embodiment of the present invention, the generation of candidate motion primitives is combined with... Figure 2 As shown, the state of the vehicle body in the inertial coordinate system s Represented as s =( x 0, y 0, φ 0), of which x 0 and y 0 represents the two-dimensional coordinate of the vehicle's center. φ 0 represents the vehicle's heading angle. v 0 represents the center velocity of the vehicle. ω0 represents the angular velocity of the vehicle body around the instantaneous turning center. In a normal gait, the robot cannot perform lateral translation, meaning the lateral velocity of the wheels is 0. Combining the vehicle's kinematic equations and geometric relationships, a simple kinematic model of the vehicle body can be easily derived: ; Among them, the motion control quantity is ( v 0, ω 0) T The vehicle control commands received by the vehicle body during the path following control process. c 0=( v 0, ω 0) T .

[0079] In this embodiment, only the most efficient Ackerman steering movement mode of the planetary rover is considered. In this mode, the middle wheel deflection angle is always 0, and the rear wheel deflection angle, in conjunction with the front wheel deflection angle, is generated based on the vehicle's parameters. Simplifying the Ackerman steering planetary rover into an equivalent bicycle model, when performing path following based on a pure tracking algorithm, the simple kinematic model can be further transformed based on geometric relationships into: ; in, δ This refers to the front wheel deflection angle, numerically... The vehicle body state variables are ( x 0, y 0, φ 0) T The motion control quantity is ( v 0, δ ) T .

[0080] For egocentric local planning problems, a robot-centric coordinate system can be used, which makes the rover's state space ( x 0, y 0, φ 0) T This is simplified to a fixed value (0, 0, 0). Since the path tracking part uses steering angle control to track the target path given the vehicle's speed, only the discretized control quantity is needed. δ Generate motion primitives. Based on the control range of the planetary rover's steering motor, the steering angle is... δ Discretize into several parts, δ Motion primitives are generated through simulation using a forward kinematics model, with the length of each motion trajectory primitive set to a multiple of the planning grid resolution.

[0081] For the training sample set, the first consideration is the acquisition of rover movement data based on ground geometry and mechanical properties. Specifically, to enable the model to capture the comprehensive impact of terrain geometry and physical properties on rover movement, movement data under various terrain conditions needs to be collected. Therefore, terrains with different combinations of geometry and physical properties are generated to collect more diverse movement data in Gazebo, facilitating subsequent model learning. Regarding terrain generation, to cover as many terrain feature combinations as possible, typical terrain geometries are selected, including flat land, slopes of varying gradients, rugged terrain, and obstacle terrain. In terms of physical properties, each terrain block is assigned a fixed block number (id) during the simulation initialization phase. At the initial stage of each data acquisition segment, parameters for each terrain block are randomized. Within their value range, the main parameters determining ground friction and bearing capacity are randomly selected. The specific randomization range and resolution of each main parameter are shown in Table 1 (settings for different terrain parameters) and Table 2 (parameter randomization settings).

[0082] Table 1. Setting Table for Different Terrain Parameters

[0083] Table 2 Parameter Randomization Settings Table

[0084] Referring to Tables 1 and 2 above, it should be noted that in Tables 1 and 2, the randomization range or randomization resolution of some parameters is marked with "-". In Table 1, "-" indicates that the parameter has not undergone randomization processing, meaning its value is fixed at the default value. For example, parameter The default value is 2350, and both the randomization range and randomization resolution are "-", indicating that during the data acquisition process, The value is always fixed at 2350 and does not change randomly. Similarly, the parameter... c The same applies to K; their values ​​are fixed at 462.3 and 0.015 respectively, without randomization. The initial positions of the parameters are shown in Table 2. and initial position The randomization resolution is marked with "-", indicating that the randomization process for this parameter does not involve resolution limitations. Specifically, the initial position The value range is [-38, 38], and the initial position is... y i The values ​​of are in the range [-38, 38], but their randomization resolution is not particularly limited, meaning that randomization can be performed with arbitrary precision within their range. Initial heading angle The randomization range is [-180, 180], with a resolution of 18. However, the randomization process of this parameter can also be unrestricted by the resolution and can be freely randomized within its value range.

[0085] Combination Figure 3 As shown, each complete simulation segment used for data acquisition consists of two processes (a stationary segment and a moving segment) and three key time points. Composition. Each simulation begins t start When the initial position is 0, the rover initializes at a random location on the map. The x and y coordinates of this initial position are obtained through random sampling within the map area, while the z coordinate is determined by the terrain height at that location. The rover's initial attitude angle... s In the middle, heading angle ψ Random sampling within the range of [-180, 180], roll angle α and pitch angle β All values ​​are initialized to 0. After the robot is initialized, a command is sent to make it stop in place and remain stationary for a period of time ∆. With the robot initially stable on the ground, its control command remains at (0, 0) for the duration of this period. If the simulation terminates prematurely during the stationary phase due to tipping over, excessive pitch, or other reasons, it indicates that the initial position is invalid. The initial state of the robot in the configuration space S is denoted as... ,Right now: ; in, The center position of the robot, Let be the heading angle of the robot in the world coordinate system.

[0086] After the static phase ends, if the robot can return to its initial state... s Once the system is stable and at rest, it enters the moving phase. At the start of the moving phase, a motion primitive is randomly sampled from the motion primitive library. m i = T ,in exist Random sampling within the range The randomization parameters are chosen to be 1 or -1 with a certain random probability. A summary of all randomization parameters is as follows: Figure 3 As shown. The path P (within a certain movement length l) is derived from this motion primitive based on robot kinematics. s , m The system follows the path using a pure tracking method and outputs vehicle control commands. The overall vehicle speed v Given in advance, ω The angular velocity around the instantaneous center is given. The planetary rover follows the path along path P( s , mThe rover moves until the simulation ends, and at the end of the simulation, the relevant information of the rover's movement for that segment is recorded.

[0087] The simulation will terminate if any of the following conditions are met: falling, collision, excessive wheel slippage, excessive wheel sinking, or exceeding the maximum time limit for each simulation segment. The movement was terminated due to failure, either exceeding the simulation map's range or reaching the path's end. Specifically, iteration failure, collision failure, excessive wheel slippage rate, excessive wheel sinking, and exceeding the simulation time all indicate that the movement segment failed. The maximum allowable wheel slippage rate is... The maximum allowable wheel sinkage is Termination due to exceeding the simulation range is a limitation of the simulation itself, and the corresponding result is not recorded; reaching the end of the path indicates that the movement of this segment has terminated successfully.

[0088] When the simulation terminates, the recorded information related to the rover's movement during that segment includes the rover's initial pose. Related map tile information Planetary Rover Motion Element The cost of completing this movement Among them, the associated map tile information Represented as: ; in, Elevation map tiles are used to represent the geometric information of the terrain. Represents a stiffness map tile. The first part represents the friction map tile, and the latter two parts are used to represent the physical information of the terrain. These terrain tiles are all centered on the robot, and the robot moves within them... State is recorded. This includes the follow motion primitive. The cost of moving Represented as: ; in, t i This indicates the time taken for that movement, i.e. , e i The energy expended is represented by the sum of the integrals of the torques at each joint and the time. Indicates whether the target point was successfully reached; 1 indicates successful completion of the journey. The movement is 0, indicating failure. If the simulation terminates prematurely during the stationary phase without random sampling of motion, the time and energy spent on randomly sampling any motion primitive will be recorded as 0 when saving the data.

[0089] In this optional embodiment, a rich training sample set is generated by collecting a large number of multimodal terrain maps, current states, candidate motion primitives, and corresponding actual time costs, energy costs, and hazard labels under different terrains in a simulated star catalog environment. This diverse data acquisition method, combined with terrain map random rotation processing, altitude drift processing, and noise perturbation processing, not only increases the diversity of data but also makes the model more robust to measurement errors or map reconstruction errors in real-world applications. Simultaneously, a multi-task loss function is used to train the initial convolutional neural network, enabling the model to simultaneously learn to predict time costs, energy costs, and hazard levels. This multi-task learning method not only improves the prediction accuracy of each task but also provides a more comprehensive decision-making basis for path planning. With the trained accessibility prediction model, the rover can predict the cost of candidate motion primitives before entering the target terrain, thereby selecting the optimal path. Furthermore, by comprehensively considering the costs of time, energy, and hazard levels, the model can provide a more comprehensive decision-making basis for path planning. This not only helps select the safest path but also optimizes path planning to reduce energy consumption and time costs, improving the overall efficiency of the mission. For example, in resource-constrained tasks, the model can prioritize paths with lower energy consumption by increasing the weight of energy cost; in time-critical tasks, the model can select the shortest path by increasing the weight of time cost. This flexible weighting mechanism enables path planning to more accurately meet the task's optimization objectives, reducing task risks and resource consumption.

[0090] In summary, this embodiment, by constructing a accessibility prediction model, not only improves the model's generalization ability and robustness, but also significantly enhances the accuracy and efficiency of path planning, reduces task risks and resource consumption, and strengthens the model's scalability and adaptability.

[0091] Optionally, the step of inputting the enhanced training sample set into the initial convolutional neural network and training the initial convolutional neural network using a multi-task loss function includes: The multimodal terrain map, current state, and candidate motion primitives in the enhanced training sample set are used as the input to the initial convolutional neural network, and the actual time cost, actual energy cost, and actual hazard labels in the enhanced training sample set are used as the supervision targets. A multi-task loss function is constructed, which is a weighted sum of time cost loss, energy cost loss, and danger level loss. The time cost loss is the mean square error between the predicted time cost and the actual time cost; the energy cost loss is the mean square error between the predicted energy cost and the actual energy cost; and the danger level loss is the cross-entropy loss or mean square error between the predicted danger level value and the actual danger label. The parameters of the initial convolutional neural network are iteratively optimized using the backpropagation algorithm, and the training of the initial convolutional neural network is completed when the loss value determined by the multi-task loss function no longer decreases after a preset number of consecutive cycles.

[0092] Specifically, the enhanced training sample set's multimodal terrain map, current state, and candidate motion primitives are used as input to the initial convolutional neural network. Simultaneously, the actual time cost, actual energy cost, and actual hazard labels in the sample set are used as supervision targets, providing the model with learning direction and optimization objectives. Secondly, a multi-task loss function is constructed, which is a weighted sum of time cost loss, energy cost loss, and hazard level loss. The time cost loss and energy cost loss are calculated using mean squared error, while the hazard level loss is selected based on the type of actual hazard label, choosing either cross-entropy loss or mean squared error. This embodiment comprehensively considers the prediction errors of multiple tasks, achieving comprehensive model optimization. Finally, the parameters of the initial convolutional neural network are iteratively optimized using the backpropagation algorithm. Each iteration adjusts the parameters according to the loss value determined by the multi-task loss function. When the loss value no longer decreases after a preset number of consecutive iterations, it indicates that the model has converged and training is complete, resulting in a passability prediction model that can accurately predict the movement cost of a rover on different terrains.

[0093] In a preferred embodiment of the present invention, combined with Figure 4 As shown, the deep learning model network structure for predicting the accessibility of a planetary rover consists of three parts: a motion encoder, a feature extractor, and a loss predictor. The motion encoder is responsible for processing motion primitives. m The motion primitives include possible actions the rover might take, such as steering angle and direction of movement. These motion primitives are encoded using a convolutional neural network (CNN) to extract motion-related features, which are then used for loss prediction. The feature extraction part consists of a geometric branch and a physical branch, processing the terrain height map H and physical property map P, respectively. The geometric branch processes the terrain height map, extracting geometric features of the terrain, such as slope and undulation. The physical branch processes the physical property map, extracting physical properties of the ground, such as friction coefficient and equivalent stiffness. The outputs of these two branches are combined to form a comprehensive feature representation used for subsequent loss prediction. The loss predictor receives the outputs from the motion encoder and the feature extraction part, predicting the cost for the rover to perform a specific motion primitive on a specific terrain. This part consists of multiple parallel branches, each predicting a specific cost dimension: energy cost. Time cost and the level of danger and cost The predicted costs are optimized using a multi-task loss function, which is a weighted sum of the losses across various cost dimensions. The multi-task loss function, LMSE, measures the difference between the predicted and actual costs. This function includes time cost loss, energy cost loss, and danger level loss, each being the mean squared error (MSE) or cross-entropy loss between the predicted and actual values. By minimizing this multi-task loss function, the model parameters are iteratively optimized until the loss value no longer decreases significantly for a predetermined number of consecutive iterations, at which point model training is complete. The model's final output is the predicted movement cost, including energy cost, time cost, and danger level cost, which will be used for the rover's path planning and decision support.

[0094] In a preferred embodiment of the invention, due to the robot's initial state Always at the center of the map tile, the robot's initial state. Represented as = T The corresponding motion m i Characterized as T The embeddings extracted from the elevation map, physical map, and local motion features are flattened and concatenated before being input into the loss prediction module. The loss prediction module consists of three branches: normalized energy loss, movement time, and hazard probability. The loss for each item is calculated using the root mean square error (MSE) between the true and predicted values, as follows: ; The model was trained using supervised learning. During data preprocessing, data augmentation techniques were employed to increase the diversity of the collected dataset. These techniques included random altitude drift, random noise in the terrain information, and random rotation. Random altitude drift involved adding a randomly sampled offset to the altitude matrix in the terrain information and the initial altitude value of the rover. This makes the learned representation insensitive to absolute terrain height, but only sensitive to relative terrain height. Random noise in terrain information refers only to the terrain height matrix. stiffness matrix and friction matrix Random noise, specifically salt-and-pepper noise, is added to enhance the robustness of the learned model against map reconstruction errors in practical applications. Random rotation refers to rotating both the terrain information matrix and the rover's initial pose information by a random sampling angle. αDuring the simulation phase, data is collected and divided into training, validation, and test sets in a certain proportion. An initial learning rate and optimizer are set, and iterations are performed to obtain a trained model and predict its feasibility.

[0095] In this optional embodiment, by simultaneously optimizing the prediction of time cost, energy cost, and risk level through a multi-task loss function, the model can more comprehensively learn the complex relationships between terrain features, vehicle states, and motion primitives, thereby significantly improving prediction accuracy. Data augmentation increases the diversity of training data, enabling the model to learn a wider range of terrain features and environmental changes. Iterative optimization of network parameters through backpropagation, stopping training when the loss value no longer decreases after a preset number of consecutive iterations, not only improves training efficiency but also ensures model training stability. This avoids overtraining or undertraining, allowing the model to reach optimal performance within a limited training time. The design of the multi-task loss function allows the model to simultaneously consider the three key factors of time, energy, and risk level, providing multi-dimensional decision-making basis for path planning. This enables path planning to not only select the shortest or least energy-consuming path but also effectively avoid high-risk areas, thereby improving the overall performance of path planning and the success rate of the task. By adjusting the weights of each loss term in the multi-task loss function, the model can flexibly adapt to different task requirements and optimization objectives. For example, in resource-constrained tasks, the weight of energy cost loss can be increased; in tasks with high safety requirements, the weight of risk level loss can be increased. This flexibility allows the model to be widely applied to a variety of complex star-based environments and mission scenarios.

[0096] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for predicting the passability of a planetary rover, characterized in that, include: Obtain multimodal terrain maps of the target star surface region, the current state of the rover, and candidate motion primitives; Input the multimodal terrain map, the current state, and the candidate motion primitives into the accessibility prediction model; The feature extractor of the accessibility prediction model is used to extract features from the multimodal terrain map to obtain shared features that couple terrain geometry and mechanics. The current state of the rover and the candidate motion primitives are combined and encoded using the motion encoder of the accessibility prediction model to obtain the combined encoding result. The combined encoding result and the shared feature are dimension-matched and concatenated to obtain a one-dimensional fused feature tensor. The loss predictor of the accessibility prediction model is used to predict the movement cost of the rover when executing the candidate motion primitives by performing loss prediction based on the one-dimensional fusion feature tensor.

2. The method for predicting the passability of a planetary rover according to claim 1, characterized in that, The acquisition of multimodal terrain maps of the target star surface region, the current state of the rover, and candidate motion primitives includes: Using the current position of the rover as the center, obtain a local elevation map, a ground friction characteristic map, and a ground softness characteristic map within a preset range of the rover; The local elevation map, the ground friction characteristic map, and the ground softness characteristic map are stacked by channels to generate the multimodal terrain map; Obtain the x-coordinate and y-coordinate of the center of the planetary rover in a preset coordinate system, as well as the heading angle of the longitudinal axis of the planetary rover relative to the y-axis of the preset coordinate system. The current state is formed based on the horizontal coordinate, the vertical coordinate, and the heading angle; The candidate motion primitives are generated based on the preset steering angle and preset direction of movement of the planetary rover.

3. The method for predicting the passability of a planetary rover according to claim 2, characterized in that, The feature extractor of the accessibility prediction model extracts features from the multimodal terrain map to obtain shared features that couple terrain geometry and mechanics, including: The local elevation map, the ground friction characteristic map, and the ground softness characteristic map from the multimodal topographic map are respectively input into the multi-channel of the feature extractor; Through the convolutional structure of the feature extraction layer, the multi-channel input terrain data is subjected to layer-by-layer feature extraction to obtain the geometric features in the local elevation map, the mechanical property distribution features in the ground friction characteristic map and the ground softness characteristic map; The feature fusion module of the feature extraction layer couples the geometric features with the mechanical property distribution features across channels to generate the shared features.

4. The method for predicting the passability of a planetary rover according to claim 2, characterized in that, The motion encoder, through the accessibility prediction model, combines and encodes the current state of the rover with the candidate motion primitives to obtain a combined encoding result, including: The horizontal coordinate, vertical coordinate, and heading angle in the current state are combined with the preset turning angle and preset movement direction in the candidate motion primitive to form a motion-state joint parameter set; The motion-state joint parameter set is numerically mapped and dimensionally transformed by the fully connected layer of the motion encoder to obtain the feature vector of the motion-state joint parameter set. The feature vector is normalized to obtain the combined encoding result.

5. The method for predicting the passability of a planetary rover according to claim 1, characterized in that, The step of performing dimensional matching and concatenation of the combined encoding result and the shared features to obtain a one-dimensional fused feature tensor includes: The shared features are flattened to convert the feature matrix of the shared features into a one-dimensional feature vector; The combined encoding result and the one-dimensional feature vector are concatenated in a first preset order to obtain the one-dimensional fused feature tensor.

6. The method for predicting the passability of a planetary rover according to claim 1, characterized in that, The loss predictor of the accessibility prediction model performs loss prediction based on the one-dimensional fused feature tensor to obtain the movement cost of the rover when executing the candidate motion primitives, including: The one-dimensional fusion feature tensor is input into the shared fully connected layer of the loss predictor to extract high-dimensional fusion features; The high-dimensional fusion features are respectively input into multiple parallel branches of the loss predictor to perform loss prediction, and the predicted value corresponding to each parallel branch is obtained. The movement cost is generated based on the predicted value of each of the parallel branches.

7. The method for predicting the passability of a planetary rover according to claim 6, characterized in that, The parallel branches include a time branch, an energy branch, and a danger level branch. The process of inputting the high-dimensional fusion features into multiple parallel branches of the loss predictor for loss prediction, and obtaining the predicted value for each parallel branch, includes: The high-dimensional fusion features are mapped through the first fully connected layer of the time branch to obtain the predicted time cost required to execute the candidate motion primitive, and the predicted time cost is used as the predicted value of the time branch. The high-dimensional fusion features are mapped through the second sub-fully connected layer of the energy branch to obtain the predicted energy cost required to execute the candidate motion primitive, and the predicted energy cost is used as the predicted value of the energy branch. The high-dimensional fusion features are mapped through the third fully connected layer of the danger level branch to obtain the predicted danger level value when executing the candidate motion primitive, and the predicted danger level value is used as the predicted value of the danger level branch.

8. The method for predicting the passability of a planetary rover according to claim 7, characterized in that, The step of generating the movement cost by concatenating or weighting the predicted values ​​of each of the parallel branches into vectors includes: The movement cost is obtained by vector combination of the predicted time cost, predicted energy cost, and predicted danger level value corresponding to the time branch, the energy branch, and the danger level branch, respectively, in a second preset order. Alternatively, corresponding weights can be assigned to the prediction time cost, the prediction energy cost, and the prediction risk level value, respectively; The movement cost is obtained by weighting and fusing the predicted time cost, the predicted energy cost, and the predicted danger level value according to their respective weights.

9. The method for predicting the passability of a planetary rover according to claim 1, characterized in that, Also includes: Construct the accessibility prediction model; The construction of the accessibility prediction model includes: In a simulated star table environment, multimodal terrain maps, current states, candidate motion primitives, and the actual time cost, energy cost, and hazard labels corresponding to the candidate motion primitives are collected under different terrain conditions to generate a training sample set. The training sample set is subjected to topographic map random rotation processing, height drift processing, and noise perturbation processing to obtain an enhanced training sample set; The enhanced training sample set is input into the initial convolutional neural network, and the initial convolutional neural network is trained using a multi-task loss function; The initial convolutional neural network after training is used as the accessibility prediction model.

10. The method for predicting the passability of a planetary rover according to claim 9, characterized in that, The step of inputting the enhanced training sample set into the initial convolutional neural network and training the initial convolutional neural network using a multi-task loss function includes: The multimodal terrain map, current state, and candidate motion primitives in the enhanced training sample set are used as the input to the initial convolutional neural network, and the actual time cost, actual energy cost, and actual hazard labels in the enhanced training sample set are used as the supervision targets. A multi-task loss function is constructed, which is a weighted sum of time cost loss, energy cost loss, and danger level loss. The time cost loss is the mean square error between the predicted time cost and the actual time cost; the energy cost loss is the mean square error between the predicted energy cost and the actual energy cost; and the danger level loss is the cross-entropy loss or mean square error between the predicted danger level value and the actual danger label. The parameters of the initial convolutional neural network are iteratively optimized using the backpropagation algorithm, and the training of the initial convolutional neural network is completed when the loss value determined by the multi-task loss function no longer decreases after a preset number of consecutive cycles.

Images