Mobile robot path planning method and device fusing wheel-ground interaction characteristics

Abstract

The application provides a mobile robot path planning method and device fusing wheel-ground interaction characteristics, and relates to the technical field of mobile robot autonomous navigation. The method comprises the following steps: initializing a global geographical environment feature map, a passing energy consumption prediction machine and a three-dimensional pose of a starting point and a target point; generating a group of discrete motion primitives based on the three-dimensional pose at the current time by using a predefined Ackerman steering geometry model, calculating all candidate sub-nodes at the next time, constructing a six-dimensional space-time feature input vector, inputting the six-dimensional space-time feature input vector into the pre-trained passing energy consumption prediction machine to obtain the predicted passing energy consumption; performing action safety constraint optimization on the action to be executed, obtaining a single-step energy consumption cost according to the three-dimensional driving distance, the passing energy consumption, the geometric feature resistance and the action safety penalty term; and selecting the optimal sub-node to generate a trajectory sequence according to the single-step energy consumption cost and a backtracking algorithm. The application can realize efficient path planning considering energy consumption optimization and driving safety.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation technology for mobile robots, and in particular to a method and apparatus for path planning of mobile robots that integrates wheel-ground interaction characteristics. Background Technology

[0002] With the rapid development of unmanned systems technology, the demand for UGVs (Unmanned Ground Vehicles) in unstructured environments such as field reconnaissance, resource exploration, disaster relief, and transportation is increasing. Unlike the structured road environment of cities, the terrain of the wilderness is extremely complex and varied. It not only presents significant geometric obstacles such as steep slopes, boulders, steps, and gullies, but also widely distributes "physical traps" with complex mechanical properties, such as seemingly flat but low-load-bearing soft mud, loose sand, or snow. These complex environmental features pose a significant challenge to the autonomous perception and planning capabilities of robots; relying solely on geometric information or simple physical models often fails to ensure the safe passage of robots.

[0003] Traditional path planning methods typically rely on 3D elevation maps constructed using LiDAR or cameras, calculating explicit geometric features such as slope and roughness to assess the cost of passage. While this method effectively avoids obvious geometric obstacles, it often leads to robot entrapment due to misjudgment when facing areas with indistinct geometric features but poor soil mechanics (such as flat mud pits). In recent years, to overcome the limitations of traditional geometric methods, data-driven energy-aware planning methods have gradually become a research hotspot. These methods attempt to utilize deep neural networks to implicitly learn the energy consumption characteristics of the terrain by analyzing robot sensor data (such as posture changes and acceleration responses), thereby achieving more intelligent path selection and improving adaptability to unknown environments.

[0004] However, existing learning-based planning methods still have many key limitations in practical applications, restricting their reliability in extremely complex environments. First, their training data often relies on simplified simulation models or very limited real-world scenarios, failing to fully cover severe nonlinear mechanical interactions such as subsidence, shear failure, and plastic deformation in real complex soil environments. This results in insufficient generalization ability of the model when facing unseen extreme conditions, leading to a significant drop in prediction accuracy. Second, these methods often suffer from dimensionality deficiencies in feature space construction. Many schemes over-rely on the robot's current propriodynamic feedback to infer environmental attributes, neglecting prior perception of subtle but crucial static geometric features on the path ahead (such as small local height discontinuities). This "shortsighted" or "blind" perception makes it difficult for the planner to avoid potential risks in advance. Furthermore, at the planning decision level, in pursuit of computational efficiency, many existing schemes oversimplify the robot's kinematic model, often treating the robot as an omnidirectionally movable point mass searching on a grid map, ignoring the robot's actual nonholonomic constraints such as Ackerman steering. This results in planned paths often containing sharp polygonal lines, making smooth tracking impossible for real vehicles. Finally, existing cost assessment mechanisms often lack a refined coupling consideration of the risks of specific actions, failing to explicitly distinguish the huge risk differences in performing different maneuvers (such as going straight in soft soil versus turning in soft soil) under different soil bearing capacity conditions, which can easily induce high-risk maneuvering behaviors that lead to mission failure.

[0005] Patent CN116147653B discloses a three-dimensional reference path planning method for autonomous vehicles. This invention constructs a layered map comprising an obstacle layer, a height layer, and a road surface semantic layer, and employs an improved A... The algorithm incorporates static stability and road surface semantic costs into the cost function to plan a reference path that meets certain off-road requirements. Although this invention introduces environmental category information through semantic segmentation, which to some extent compensates for the shortcomings of pure geometric maps, it still essentially relies on the static superposition of visual semantics and geometric slope. It lacks in-depth modeling of the complex dynamic interaction mechanism between wheels and soil (such as shear settlement and slip effect), which makes it easy to get stuck in the road surface due to misjudging the physical properties when facing soft mud or loose sand areas that are "visually flat but have extremely low physical bearing capacity." Furthermore, it does not fully consider the actual differences in the dynamic response of vehicles under complex terrain.

[0006] Patent CN115357022B discloses a method for 3D path planning of off-road vehicles based on deep reinforcement learning. This invention constructs a value-based deep convolutional neural network model, designs a reward function that comprehensively considers distance and energy consumption, and achieves off-road path planning through end-to-end training. However, it adopts an end-to-end black-box learning strategy, and the model lacks interpretability and explicit physical constraints, making it difficult to guarantee safety under extreme conditions. At the same time, this method often ignores the inherent nonholonomic kinematic constraints of Ackermann steering vehicles (such as HMMWVs, which are high-mobility multipurpose wheeled vehicles), and the generated paths are prone to curvature abrupt changes or unexecutable broken lines. Furthermore, its training environment is mostly based on simplified kinematic models and lacks the support of high-fidelity soil mechanics simulation data, resulting in insufficient generalization ability of the model in real complex soft soil interaction scenarios, making it difficult to apply directly. Summary of the Invention

[0007] To address the technical problems of existing data-driven methods, such as training sets relying heavily on simplified kinematic models or limited real-vehicle data collection, lacking characterization of extreme soil environments; missing input feature dimensions in existing models, easily leading to hidden risks of vehicles getting stuck; and poor generalization ability and prediction reliability, this invention provides a mobile robot path planning method and apparatus that integrates wheel-ground interaction characteristics. The technical solution is as follows: On the one hand, a mobile robot path planning method integrating wheel-ground interaction characteristics is provided. This method is implemented by a mobile robot path planning device and includes: S1. Initialize the offline pre-computed global geographic environment feature map, the pre-trained traffic energy consumption predictor, and the 3D pose of the robot's starting point and target point.

[0008] S2. Based on the robot's current 3D pose, generate a set of discrete motion primitives that conform to the robot's nonholonomic constraints using a predefined Ackerman steering geometry model, and calculate the position, heading, and intended action of all feasible candidate child nodes in the next moment based on the discrete motion primitives.

[0009] S3. Based on the global geographic environment feature map, a six-dimensional spatiotemporal feature input vector is constructed for each candidate child node and its corresponding intended action. The six-dimensional spatiotemporal feature input vector is then input into a pre-trained traffic energy consumption predictor to obtain the predicted traffic energy consumption.

[0010] S4. Optimize the motion safety constraints of the proposed action to obtain the motion safety penalty term. Based on the three-dimensional travel distance from the robot to the candidate sub-node, the predicted travel energy consumption, the geometric feature resistance, and the motion safety penalty term, obtain the single-step energy consumption cost of each candidate sub-node.

[0011] S5. Select the optimal child node based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm, generate the trajectory sequence, and complete the robot path planning.

[0012] Optionally, the six-dimensional spatiotemporal feature input vector in S3 includes: four-dimensional historical dynamic features and two-dimensional forward geometric features.

[0013] Among them, the four-dimensional historical dynamic features include the mean pitch angle, mean roll angle, standard deviation of pitch angle, and standard deviation of roll angle within the robot body's past preset time window, which are used to implicitly characterize the physical and mechanical properties of the soil.

[0014] Two-dimensional forward geometric features include: local roughness and local height difference of candidate child nodes read from the global geographic environment feature map, used to explicitly characterize the terrain morphology of the candidate child node location.

[0015] Optionally, the pre-training process of the traffic energy consumption predictor in S1 includes: A robot dynamics model is established in a multibody dynamics simulation engine, and a soft soil contact model is introduced to generate unstructured terrain through a randomized parameter set.

[0016] By controlling the robot to perform diverse discrete motion primitives in unstructured terrain, the robot's real physical response and energy consumption are obtained, and a high-confidence physical dataset of state-action-energy consumption is constructed.

[0017] Based on a high-confidence physical dataset of state-action-energy consumption, a deep neural network model based on bidirectional long short-term memory network and attention mechanism is trained to obtain a pre-trained travel energy consumption predictor.

[0018] Optionally, a deep neural network model based on a bidirectional long short-term memory network and an attention mechanism is trained using a high-confidence physics dataset of state-action-energy consumption, including: Based on a high-confidence physics dataset of state-action-energy consumption, four-dimensional historical dynamic features and two-dimensional forward geometric features are obtained.

[0019] The four-dimensional historical dynamic features are input into the first bidirectional long short-term memory network to obtain the forward hidden state and the backward hidden state. The forward hidden state and the backward hidden state are concatenated to obtain the soil feature encoding sequence. The two-dimensional forward geometric features are used as the query vector, and the soil feature encoding sequence is used as the key vector and value vector. The context vector containing physical perception is generated by weighted summation.

[0020] The four-dimensional historical dynamic features are input into the second bidirectional long short-term memory network to obtain the temporal hidden feature vector. The temporal hidden feature vector is used as the query vector, the context vector and the two-dimensional forward geometric features are concatenated as the key vector, the proposed action corresponding to the candidate child node is used as the value vector, and the predicted passage energy consumption is output through the fully connected layer.

[0021] Optionally, the action safety constraint optimization in S4 includes: Based on four-dimensional historical dynamics features, it identifies whether the action to be performed belongs to high-risk motion behavior. For the identified high-risk motion behavior, a nonlinear high penalty factor is applied or adaptive pruning is triggered to remove the high-risk motion behavior.

[0022] Optionally, in S5, the optimal child node is selected based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm to generate a trajectory sequence, including: A global topological potential field graph is generated using a reverse search algorithm. Based on the global topological potential field graph, the Euclidean distance and topological potential field distance from the candidate child nodes to the target point are obtained. The maximum value of the Euclidean distance and the topological potential field distance is selected as the heuristic prediction cost. The optimal child node is selected from the candidate child nodes based on the sum of the single-step energy consumption cost and the heuristic prediction cost, and a trajectory sequence is generated.

[0023] On the other hand, a mobile robot path planning device integrating wheel-ground interaction characteristics is provided. This device is applied to a mobile robot path planning method integrating wheel-ground interaction characteristics. The device includes: The initialization module is used to initialize the offline pre-computed global geographic environment feature map, the pre-trained traffic energy consumption predictor, and the 3D pose of the robot's starting and target points.

[0024] The candidate child node calculation module is used to generate a set of discrete motion primitives that conform to the robot's nonholonomic constraints based on the robot's current 3D pose using a predefined Ackerman steering geometry model, and to calculate the position, heading, and intended action of all feasible candidate child nodes in the next moment based on the discrete motion primitives.

[0025] The traffic energy consumption prediction module is used to construct a six-dimensional spatiotemporal feature input vector for each candidate child node and its corresponding intended action based on the global geographic environment feature map. The six-dimensional spatiotemporal feature input vector is then input into a pre-trained traffic energy consumption prediction machine to obtain the predicted traffic energy consumption.

[0026] The single-step energy consumption cost calculation module is used to optimize the action safety constraints of the action to be executed and obtain the action safety penalty term. Based on the three-dimensional travel distance from the robot to the candidate sub-node, the predicted travel energy consumption, the geometric feature resistance, and the action safety penalty term, the single-step energy consumption cost of each candidate sub-node is obtained.

[0027] The output module is used to select the optimal child node based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm, generate a trajectory sequence, and complete the robot path planning.

[0028] Optionally, the six-dimensional spatiotemporal feature input vector includes: four-dimensional historical dynamic features and two-dimensional forward geometric features.

[0029] Among them, the four-dimensional historical dynamic features include the mean pitch angle, mean roll angle, standard deviation of pitch angle, and standard deviation of roll angle within the robot body's past preset time window, which are used to implicitly characterize the physical and mechanical properties of the soil.

[0030] Two-dimensional forward geometric features include: local roughness and local height difference of candidate child nodes read from the global geographic environment feature map, used to explicitly characterize the terrain morphology of the candidate child node location.

[0031] Optionally, the initialization module is further used for: A robot dynamics model is established in a multibody dynamics simulation engine, and a soft soil contact model is introduced to generate unstructured terrain through a randomized parameter set.

[0032] By controlling the robot to perform diverse discrete motion primitives in unstructured terrain, the robot's real physical response and energy consumption are obtained, and a high-confidence physical dataset of state-action-energy consumption is constructed.

[0033] Based on a high-confidence physical dataset of state-action-energy consumption, a deep neural network model based on bidirectional long short-term memory network and attention mechanism is trained to obtain a pre-trained travel energy consumption predictor.

[0034] Optionally, the initialization module is further used for: Based on a high-confidence physics dataset of state-action-energy consumption, four-dimensional historical dynamic features and two-dimensional forward geometric features are obtained.

[0035] The four-dimensional historical dynamic features are input into the first bidirectional long short-term memory network to obtain the forward hidden state and the backward hidden state. The forward hidden state and the backward hidden state are concatenated to obtain the soil feature encoding sequence. The two-dimensional forward geometric features are used as the query vector, and the soil feature encoding sequence is used as the key vector and value vector. The context vector containing physical perception is generated by weighted summation.

[0036] The four-dimensional historical dynamic features are input into the second bidirectional long short-term memory network to obtain the temporal hidden feature vector. The temporal hidden feature vector is used as the query vector, the context vector and the two-dimensional forward geometric features are concatenated as the key vector, the proposed action corresponding to the candidate child node is used as the value vector, and the predicted passage energy consumption is output through the fully connected layer.

[0037] Optionally, the single-step energy consumption cost calculation module is further used for: Based on four-dimensional historical dynamics features, it identifies whether the action to be performed belongs to high-risk motion behavior. For the identified high-risk motion behavior, a nonlinear high penalty factor is applied or adaptive pruning is triggered to remove the high-risk motion behavior.

[0038] Optionally, the output module is further used for: A global topological potential field graph is generated using a reverse search algorithm. Based on the global topological potential field graph, the Euclidean distance and topological potential field distance from the candidate child nodes to the target point are obtained. The maximum value of the Euclidean distance and the topological potential field distance is selected as the heuristic prediction cost. The optimal child node is selected from the candidate child nodes based on the sum of the single-step energy consumption cost and the heuristic prediction cost, and a trajectory sequence is generated.

[0039] On the other hand, a mobile robot path planning device is provided, the mobile robot path planning device comprising: a processor; a memory storing computer-readable instructions, wherein when the computer-readable instructions are executed by the processor, any one of the above-described mobile robot path planning methods that integrate wheel-ground interaction characteristics is implemented.

[0040] On the other hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement any of the above-described mobile robot path planning methods that integrate wheel-ground interaction characteristics.

[0041] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: This invention constructs a high-fidelity physical interaction data foundation based on the SCM (Soil Contact Model), fundamentally filling the gap in prior physical data for extreme off-road conditions. Traditional data-driven methods often rely on simplified kinematic models or limited real-vehicle data collection for their training sets, lacking a characterization of extreme soil environments. This invention, relying on an advanced soft soil contact mechanics model, synchronously records the robot's state, actions, and actual energy consumption under complex soil conditions at high frequency, deeply capturing nonlinear wheel-ground interaction phenomena such as "shear subsidence" and "multipath slippage." This high-confidence dataset provides a training source for subsequent deep learning models that absolutely conforms to real physical laws, solving the technical problem of prediction distortion in existing models under extreme conditions.

[0042] This invention innovatively defines a "six-dimensional spatiotemporal feature space" for visual-tactile fusion, completely breaking through the perceptual blind spots of single feature input. Addressing the pain point of existing models lacking input feature dimensions, which easily leads to the hidden risk of vehicles getting stuck, this invention spatially aligns and stitches together a four-dimensional high-frequency dynamic sequence representing the interaction between the chassis and the ground (implicit tactile features) with a two-dimensional elevation map of the area to be planned ahead (explicit visual features). This novel feature input construction endows the trained model with a composite perceptual ability to "predict the danger of the slope ahead by combining the softness and hardness underfoot," effectively avoiding misjudging "visually flat but extremely low-load-bearing muddy ground" as a safe area.

[0043] This invention significantly improves the generalization ability and prediction reliability of predictive machine models based on input features with deep physical coupling. Because the training data of the input model is not simply a geometric spatial representation, but deeply coupled with the nonlinear interactive physical effects between the robot chassis and the complex soil environment (such as sensing unobservable parameters like soil damping and internal friction angle through equivalent perception of attitude standard deviation), the time-series deep learning model trained on this highly physically coupled dataset can truly learn the mechanical mechanisms of wheel-ground interaction. Therefore, even when facing unseen complex soft soil traps, it can still perform high-fidelity energy consumption cost estimations, ensuring absolute safety in risk decision-making. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a mobile robot path planning method that integrates wheel-ground interaction characteristics, provided by an embodiment of the present invention. Figure 2 This is a diagram of a deep neural network model based on Bi-LSTM and Attention mechanism provided in an embodiment of the present invention; Figure 3 This is a block diagram of a mobile robot path planning device that integrates wheel-ground interaction characteristics, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of a mobile robot path planning device provided in an embodiment of the present invention. Detailed Implementation

[0046] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0047] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0048] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0049] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0050] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0051] This invention provides a mobile robot path planning method that integrates wheel-ground interaction characteristics. This method can be implemented by a mobile robot path planning device, which can be a terminal or a server. Figure 1 The flowchart shown is for a mobile robot path planning method that integrates wheel-ground interaction characteristics. The processing flow of this method may include the following steps: S1. Initialize the offline pre-computed global geographic environment feature map, the pre-trained traffic energy consumption predictor, and the 3D pose of the robot's starting point and target point.

[0052] One feasible implementation involves initializing the key variables and prior knowledge required for planning. This includes loading an offline pre-computed global geographic environment feature map (containing local height difference (LHD) and surface roughness information) and loading a deep neural network model trained on high-fidelity physical simulation data. This deep neural network model is a fusion model of Bi-LSTM (Bi-directional Long Short-Term Memory) and an attention mechanism, as shown in the figure below. Figure 2 As shown, the three-dimensional pose states of the robot's starting point and target point are displayed. The corresponding grid index is included in the search set.

[0053] Optionally, the pre-training process of the traffic energy consumption predictor in S1 may include the following steps S11-S13: S11. Establish a robot dynamics model in the multibody dynamics simulation engine and introduce a soft soil contact model to generate unstructured terrain through a randomized parameter set.

[0054] In one feasible implementation, in order to completely overcome the fatal defects of existing data-driven planning methods such as "training data distortion", "missing feature dimensions" and "poor generalization ability", this invention proposes a method for constructing a visual-tactile multimodal dataset based on high-fidelity soil mechanics simulation, and defines a novel six-dimensional spatiotemporal feature space.

[0055] Specifically, this invention proposes a high-fidelity physical interaction data acquisition method based on SCM (Solid Motion Computation). Specifically, this invention first establishes an accurate robot dynamics model in a multibody dynamics simulation engine (such as Pychrono) and then introduces an advanced soft soil contact mechanics model using SCM. This is achieved by randomizing the Bekker parameter set. (Subsidence index, cohesion, internal friction angle, etc.) The system generates highly challenging unstructured terrains covering hard surfaces, soft sand, and sticky mud.

[0056] S12. By controlling the robot to execute diverse discrete motion primitives in unstructured terrain, the robot's real physical response and energy consumption are obtained, and a high-confidence physical dataset of state-action-energy consumption is constructed.

[0057] In one feasible implementation, by controlling a virtual robot to execute diverse discrete motion primitives in the aforementioned extreme terrain, the system synchronously records the robot's actual physical response and energy consumption at high frequency. This resulted in a high-confidence physical dataset of "state-action-energy consumption". This data base deeply captures the nonlinear mechanical energy consumption in real off-road environments such as "shear subsidence" and "multipath slip ratio", fundamentally filling the prior data gap in extreme physical trap environments.

[0058] S13. Based on the high-confidence physical dataset of state-action-energy consumption, train the deep neural network model based on bidirectional long short-term memory network and attention mechanism to obtain a pre-trained traffic energy consumption predictor.

[0059] Optionally, step S13 above specifically includes: S131. Based on the high-confidence physical dataset of state-action-energy consumption, obtain the four-dimensional historical dynamic features and the two-dimensional forward geometric features.

[0060] In one feasible implementation, this invention proposes an innovative definition of a "six-dimensional spatiotemporal feature space" for visual-tactile fusion. Specifically, this invention innovatively defines a six-dimensional spatiotemporal feature vector to comprehensively characterize the robot's current physical state and the environmental attributes in front of it. This feature vector consists of four-dimensional historical dynamic features and two-dimensional forward geometric features: ① Four-dimensional historical dynamic characteristics: used to implicitly characterize the physical and mechanical properties of soil. Capturing the current moment. Previous fixed time window For robot pose data within 1.0 second (e.g.), the following statistics are calculated: a. Mean pitch angle With the mean roll angle : Characterizes the robot's average posture tendency.

[0061] b. Standard deviation of pitch angle : Characterizes longitudinal vibration intensity and is used to identify the viscoelasticity and damping properties of soil (such as low-frequency, high-damping vibrations in soft soil).

[0062] c. Standard deviation of roll angle : Characterizes the intensity of lateral swaying and reflects the irregular disturbance of the terrain.

[0063] The aforementioned four-dimensional time-series statistics are not simply geometric spatial attitude measurements, but rather deeply coupled nonlinear interactive physical effects between the robot chassis and the complex soil environment. Specifically, when the robot enters a low-load-bearing soft soil area, the shear settlement of the wheels will directly affect the mean pitch and roll angles. , Steady-state deviation occurs; simultaneously, the high damping characteristics of soft soil and the non-periodic jolts caused by wheel slippage will significantly alter the standard deviation of the attitude sequence. , Therefore, by extracting high-frequency temporal attitude features, this invention effectively achieves 'implicit tactile' perception of unobservable soil mechanical parameters (such as subsidence index, internal friction angle, etc.) at the bottom layer.

[0064] ② Two-dimensional forward geometric features: used to explicitly characterize the terrain features of the location to be planned. Based on the airborne local elevation map, at the node to be planned... Calculation at point: a. Local roughness The degree of surface gravel is characterized by calculating the normal vector dispersion of a fitted local microplane.

[0065] b. Local height difference (LHD): Calculates the elevation difference between the highest and lowest points within a local window, representing geometric obstacles such as steps and potholes.

[0066] This 6D feature set gives the robot human-like intuition to "predict the danger of the slope ahead by combining the softness and hardness of the ground under its feet," enabling subsequent models to make high-fidelity composite predictions.

[0067] S132. Input the four-dimensional historical dynamic features into the first bidirectional long short-term memory network to obtain the forward hidden state and the backward hidden state. Concatenate the forward hidden state and the backward hidden state to obtain the soil feature encoding sequence. Use the two-dimensional forward geometric features as the query vector and the soil feature encoding sequence as the key vector and value vector. Generate a context vector containing physical perception by weighted summation.

[0068] In one feasible implementation, the symbolic definition of a multidimensional feature vector is: a set of historical dynamic sequences. With the front geometry-action query vector . Indicates the first The robot's state vector matrix at time t. A static attribute matrix representing the location to be planned.

[0069] Furthermore, soil physical property encoding based on Bi-LSTM: A bidirectional long short-term memory network is used to extract time-dependent features from historical sequences to implicitly encode soil rheological and mechanical properties (such as damping and stiffness). The sequences are then processed... Input the Bi-LSTM layer and compute the forward hidden state. and hidden state : (1) (2) By concatenating the hidden states in the two directions, we obtain the first... The comprehensive context vector at each moment Forming soil feature coding sequences .

[0070] Furthermore, feature fusion based on an attention mechanism: To quantify the influence weight of historical soil characteristics on energy consumption under the current geometry, an attention mechanism is introduced. The geometric query vector... Mapping to a query encodes soil features. The mapping is done as keys and values. A weighted summation is then used to generate the final context vector containing physical awareness. This vector It integrates historical physical features most relevant to the current terrain, serving as the core intermediate input for subsequent energy consumption prediction and realizing the biomimetic perception logic of "judging road conditions based on foot feel".

[0071] The attention layer in this step is used for geometric-physical feature fusion. Query is the current geometric query, Key is the historical soil feature encoding, and Value is the historical soil feature encoding, which is used to resolve the influence of the current geometric terrain on the historical soil.

[0072] Furthermore, energy consumption prediction and loss function optimization: using the context vector Compared with the original query vector After concatenation, the data is input into a fully connected layer (MLP) and the predicted energy consumption coefficient is output. Regularization terms are added to prevent overfitting.

[0073] S133. Input the four-dimensional historical dynamic features into the second bidirectional long short-term memory network to obtain the temporal hidden feature vector. Use the temporal hidden feature vector as the query vector, concatenate the context vector and the two-dimensional forward geometric features as the key vector, use the proposed action corresponding to the candidate child node as the value vector, and output the predicted passage energy consumption through the fully connected layer.

[0074] In one feasible implementation, in order to establish a nonlinear mapping from six-dimensional features to travel energy consumption, a deep neural network model based on Bi-LSTM and Attention mechanism is constructed.

[0075] Specifically, firstly, the four-dimensional historical dynamic sequence is input into a bidirectional long short-term memory network to extract temporal latent feature vectors. This vector encodes the current soil bearing capacity information. Subsequently, an attention mechanism is introduced to... The query vector uses the two-dimensional geometric features of the point to be planned as the key vector, and the proposed actions corresponding to the candidate child nodes as the value vectors for feature fusion. In this process, the discrete proposed actions are pre-converted into continuous high-dimensional vectors as the value vectors. After weighted summation via an attention mechanism, the output context vector not only contains the intrinsic information of the proposed actions but also deeply couples the weights of historical dynamic states and the terrain ahead, representing the comprehensive action cost features under a specific physical environment. Finally, a nonlinear regression mapping is constructed through a fully connected layer (MLP) to reduce the dimensionality of this high-dimensional comprehensive action cost feature and output the predicted continuous passage energy consumption coefficient. The model uses mean squared error as the loss function and is trained under supervised supervision on the generated dataset. The trained model is then called an "online physics cost predictor," capable of responding in real time during planning.

[0076] The attention layer in this step is used for action energy consumption prediction. Query is the historical dynamic features, Key is the geometric features of the point to be planned, and Value is the action to be executed. It is used to solve the problem of historical physical state → future action energy consumption.

[0077] This invention fuses historical soil features with current topographic geometry features to output a fused feature vector. , fuse feature vectors The data is combined with the original terrain geometry features and used as input for the Attention mechanism. This data, along with historical dynamic features and the intended action, is then fused together to output the final product. This final product is then fed into a fully connected layer, directly outputting the final energy consumption coefficient. .

[0078] S2. Based on the robot's current 3D pose, generate a set of discrete motion primitives that conform to the robot's nonholonomic constraints using a predefined Ackerman steering geometry model, and calculate the position, heading, and intended action of all feasible candidate child nodes in the next moment based on the discrete motion primitives.

[0079] In one feasible implementation, the present invention utilizes an improved Hybrid A The algorithm performs real-time path search.

[0080] Specifically, based on the robot's current three-dimensional pose, the algorithm abandons the traditional omnidirectional movement assumption and uses a predefined Ackerman turning geometry model to generate a set of discrete motion primitives that conform to the robot's nonholonomic constraints (such as sharp left turn, straight movement, and gentle right turn). Based on this, it calculates the positions and headings of all feasible candidate child nodes in the next moment.

[0081] S3. Based on the global geographic environment feature map, a six-dimensional spatiotemporal feature input vector is constructed for each candidate child node and its corresponding intended action. The six-dimensional spatiotemporal feature input vector is then input into a pre-trained traffic energy consumption predictor to obtain the predicted traffic energy consumption.

[0082] In one feasible implementation, the present invention introduces a key “online physical cost predictor” step during the planning process to evaluate path points in order to deeply characterize the complex physical geometry of the soil.

[0083] Specifically, for each candidate child node and its corresponding intended action, the algorithm constructs a six-dimensional spatiotemporal feature input vector in real time: it includes a four-dimensional high-frequency dynamic response sequence extracted from a window of time in the past of the robot body (mean pitch angle, mean roll angle and their respective standard deviations, used to implicitly characterize physical properties such as soil damping and bearing capacity), and two-dimensional static geometric information of the planned position read from the feature map (local height difference and roughness, used to explicitly characterize terrain geometric properties).

[0084] S4. Optimize the motion safety constraints of the proposed action to obtain the motion safety penalty term. Based on the three-dimensional travel distance from the robot to the candidate sub-node, the predicted travel energy consumption, the geometric feature resistance, and the motion safety penalty term, obtain the single-step energy consumption cost of each candidate sub-node.

[0085] In one feasible implementation, in order to address the risk of vehicles getting stuck in complex soft soil terrain, this invention optimizes the physical cost sequence predicted by the model by performing action safety constraints.

[0086] Specifically, this optimization step is based on the judgment of soil mechanical properties implied by model inference. It applies a nonlinear high penalty factor to driving behaviors identified as high-risk (such as making large-angle turns in soft soil with low bearing capacity), thereby forcibly avoiding risky paths that may lead to shear settlement at the planning level and ensuring the safety of the control strategy.

[0087] S5. Select the optimal child node based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm, generate the trajectory sequence, and complete the robot path planning.

[0088] Optionally, in S5, the optimal child node is selected based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm to generate a trajectory sequence, including: A global topological potential field graph is generated using a reverse search algorithm. Based on the global topological potential field graph, the Euclidean distance and topological potential field distance from the candidate child nodes to the target point are obtained. The maximum value of the Euclidean distance and the topological potential field distance is selected as the heuristic prediction cost. The optimal child node is selected from the candidate child nodes based on the sum of the single-step energy consumption cost and the heuristic prediction cost, and a trajectory sequence is generated.

[0089] In one feasible implementation, during the path tracking execution phase, the present invention generates a complete trajectory sequence containing continuous position coordinates, heading angles, and corresponding control action commands based on optimal search tree backtracking. The underlying path tracking controller (such as the Pure Pursuit controller) receives this trajectory sequence and executes the robot's longitudinal speed and lateral steering control step by step until the robot finally and safely reaches the set target point.

[0090] This invention proposes an ultra-fast online optimization based on feedforward cost tensor and Ackerman kinematics. Specifically, this stage aims to completely decouple the computational conflict between "deep learning inference" and "high-frequency path search," achieving millisecond-level planning that integrates physical constraints through an architecture of "pre-computation combined with ultra-fast table lookup." Before the path planner starts, the system constructs a locally sliding look-ahead map that dynamically updates as the robot moves, based on the effective look-ahead field of view of onboard sensors (such as LiDAR). The system packages all the 'spatial mesh pose-execution action' combinations within this locally sliding map and inputs them all at once into the previously trained physics prediction machine model, generating a three-dimensional physical cost tensor covering the safe distance in front of the robot through parallel inference. As the robot moves, the system updates the physical cost value at the edges of this tensor incrementally in milliseconds.

[0091] In the subsequent high-frequency node expansion and optimization loop, this invention abandons the traditional assumption of omnidirectional mass movement. Based on the robot's wheelbase and maximum steering angle limitations, the system calls predefined Ackerman discrete control primitives (such as straight-line, slow-turn, and sharp-turn actions) to generate candidate child nodes for the next moment, ensuring the geometric continuity and underlying physical controllability of the generated path. Simultaneously, to address the severe undulations of the terrain, the algorithm extracts the elevation difference between nodes and uses spatial geometric relationships to restore the step distance in the two-dimensional plane to the actual three-dimensional travel distance, effectively correcting the mileage prediction error caused by neglecting gravitational potential energy.

[0092] In the single-step cost calculation stage, the planner no longer calls the neural network, but directly uses the coordinates of the current candidate grid and the index of the action to be executed as the key value, and extracts the corresponding travel energy consumption coefficient in constant time from the pre-calculated three-dimensional physical cost tensor. To prevent the robot from performing large-angle turns in soft soil and causing shear subsidence, the system introduces a dynamic penalty mechanism based on physical perception: when the lookup result shows that the current area is in a low-bearing-capacity soft soil environment, the algorithm applies a nonlinear high penalty to high-curvature actions, and may even trigger adaptive pruning to directly remove the action branch. Subsequently, the algorithm performs a comprehensive multiplication and addition operation on the actual three-dimensional travel distance, the physical resistance coefficient obtained by the predictor from the lookup table, the geometric feature resistance, and the action safety penalty term to obtain a single-step energy consumption cost that is extremely close to the real physical interaction. Among them, the geometric feature resistance corresponds to the "two-dimensional forward geometric features" (i.e., local roughness and local height difference) defined above in this invention. This geometric feature resistance is mainly composed of two explicit spatial geometric physical quantities: roughness damping and height difference / gravity resistance (slope resistance).

[0093] Finally, to improve the macroscopic search guidance and avoid local deadlocks in large physical traps or U-shaped obstacles, this invention generates a global topological potential field map using a reverse search algorithm before formal planning. During node evaluation, the algorithm dynamically compares the "pure Euclidean distance" and "topological potential field distance" from the current node to the destination, taking the maximum value as the tightest lower bound heuristically estimated cost. The planner continuously selects the node with the smallest sum of cumulative actual cost and heuristically estimated cost for expansion until it enters the target point's gravity threshold circle, outputting a final smooth and absolutely anti-sinking safe off-road trajectory.

[0094] This invention proposes a path planning method for mobile robots that integrates wheel-ground interaction characteristics, applicable to complex and unknown unstructured outdoor environments. It employs a two-stage architecture of "offline physical learning - online perception planning," combined with an improved A* algorithm based on Ackerman motion primitives. Algorithms and temporal deep learning physical models enable efficient path planning that balances energy consumption optimization and driving safety. Details are as follows:

[0095] In this invention, a high-fidelity physical interaction data base based on SCM is constructed, fundamentally filling the gap in prior physical data for extreme off-road conditions. Traditional data-driven methods often rely on simplified kinematic models or limited real-vehicle data collection for their training sets, lacking a characterization of extreme soil environments. This invention, based on an advanced soft soil contact mechanics model, synchronously records the robot's state, actions, and actual energy consumption under complex soil conditions at high frequency, deeply capturing nonlinear wheel-ground interaction phenomena such as "shear subsidence" and "multipath slippage." This high-confidence dataset provides a training source for subsequent deep learning models that absolutely conforms to real physical laws, solving the technical problem of prediction distortion in existing models under extreme conditions.

[0096] This invention innovatively defines a "six-dimensional spatiotemporal feature space" for visual-tactile fusion, completely breaking through the perceptual blind spots of single feature input. Addressing the pain point of existing models lacking input feature dimensions, which easily leads to the hidden risk of vehicles getting stuck, this invention spatially aligns and stitches together a four-dimensional high-frequency dynamic sequence representing the interaction between the chassis and the ground (implicit terrain features) with a two-dimensional elevation map of the area to be planned ahead (explicit terrain features). This novel feature input construction endows the trained model with a composite perceptual ability to "predict the danger of the slope ahead by combining the softness and hardness underfoot," effectively avoiding misjudging "visually flat but extremely low-load-bearing muddy ground" as a safe area.

[0097] This invention significantly improves the generalization ability and prediction reliability of predictive machine models based on input features with deep physical coupling. Because the training data of the input model is not simply a geometric spatial representation, but deeply coupled with the nonlinear interactive physical effects between the robot chassis and the complex soil environment (such as sensing unobservable parameters like soil damping and internal friction angle through equivalent perception of attitude standard deviation), the time-series deep learning model trained on this highly physically coupled dataset can truly learn the mechanical mechanisms of wheel-ground interaction. Therefore, even when facing unseen complex soft soil traps, it can still perform high-fidelity energy consumption cost estimations, ensuring absolute safety in risk decision-making.

[0098] Figure 3 This is a block diagram illustrating a mobile robot path planning device that integrates wheel-ground interaction characteristics according to an exemplary embodiment. The device is used in a mobile robot path planning method that integrates wheel-ground interaction characteristics. (Refer to...) Figure 3 The device includes an initialization module 310, a candidate child node calculation module 320, a passage energy consumption prediction module 330, a single-step energy consumption cost calculation module 340, and an output module 350. Wherein: The initialization module 310 is used to initialize the offline pre-computed global geographic environment feature map, the pre-trained traffic energy consumption predictor, and the three-dimensional pose of the robot's starting point and target point.

[0099] The candidate child node calculation module 320 is used to generate a set of discrete motion primitives that conform to the nonholonomic constraints of the robot based on the robot's current three-dimensional pose using a predefined Ackerman steering geometry model, and to calculate the position, heading and intended action of all feasible candidate child nodes in the next moment based on the discrete motion primitives.

[0100] The traffic energy consumption prediction module 330 is used to construct a six-dimensional spatiotemporal feature input vector for each candidate child node and its corresponding intended action based on the global geographic environment feature map. The six-dimensional spatiotemporal feature input vector is then input into a pre-trained traffic energy consumption predictor to obtain the predicted traffic energy consumption.

[0101] The single-step energy consumption cost calculation module 340 is used to optimize the action safety constraints of the action to be executed, and obtain the action safety penalty term. Based on the three-dimensional travel distance from the robot to the candidate sub-node, the predicted travel energy consumption, the geometric feature resistance, and the action safety penalty term, the single-step energy consumption cost of each candidate sub-node is obtained.

[0102] The output module 350 is used to select the optimal child node based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm, generate a trajectory sequence, and complete the robot path planning.

[0103] In this invention, a high-fidelity physical interaction data base based on SCM is constructed, fundamentally filling the gap in prior physical data for extreme off-road conditions. Traditional data-driven methods often rely on simplified kinematic models or limited real-vehicle data collection for their training sets, lacking a characterization of extreme soil environments. This invention, based on an advanced soft soil contact mechanics model, synchronously records the robot's state, actions, and actual energy consumption under complex soil conditions at high frequency, deeply capturing nonlinear wheel-ground interaction phenomena such as "shear subsidence" and "multipath slippage." This high-confidence dataset provides a training source for subsequent deep learning models that absolutely conforms to real physical laws, solving the technical problem of prediction distortion in existing models under extreme conditions.

[0104] This invention innovatively defines a "six-dimensional spatiotemporal feature space" for visual-tactile fusion, completely breaking through the perceptual blind spots of single feature input. Addressing the pain point of existing models lacking input feature dimensions, which easily leads to the hidden risk of vehicles getting stuck, this invention spatially aligns and stitches together a four-dimensional high-frequency dynamic sequence representing the interaction between the chassis and the ground (implicit tactile features) with a two-dimensional elevation map of the area to be planned ahead (explicit visual features). This novel feature input construction endows the trained model with a composite perceptual ability to "predict the danger of the slope ahead by combining the softness and hardness underfoot," effectively avoiding misjudging "visually flat but extremely low-load-bearing muddy ground" as a safe area.

[0105] This invention significantly improves the generalization ability and prediction reliability of predictive machine models based on input features with deep physical coupling. Because the training data of the input model is not simply a geometric spatial representation, but deeply coupled with the nonlinear interactive physical effects between the robot chassis and the complex soil environment (such as sensing unobservable parameters like soil damping and internal friction angle through equivalent perception of attitude standard deviation), the time-series deep learning model trained on this highly physically coupled dataset can truly learn the mechanical mechanisms of wheel-ground interaction. Therefore, even when facing unseen complex soft soil traps, it can still perform high-fidelity energy consumption cost estimations, ensuring absolute safety in risk decision-making.

[0106] Figure 4 This is a structural schematic diagram of a mobile robot path planning device provided in an embodiment of the present invention, as shown below. Figure 4 As shown, the mobile robot path planning device may include the above-mentioned Figure 3 The illustrated mobile robot path planning device integrates wheel-ground interaction characteristics. Optionally, the mobile robot path planning device 410 may include a first processor 2001.

[0107] Optionally, the mobile robot path planning device 410 may also include a memory 2002 and a transceiver 2003.

[0108] The first processor 2001, memory 2002, and transceiver 2003 can be connected via a communication bus.

[0109] The following is combined with Figure 4 A detailed description of each component of the mobile robot path planning device 410 is provided below: The first processor 2001 is the control center of the mobile robot path planning device 410. It can be a single processor or a collective term for multiple processing elements. For example, the first processor 2001 can be one or more central processing units (CPUs), application-specific integrated circuits (ASICs), or one or more integrated circuits configured to implement embodiments of the present invention, such as one or more digital signal processors (DSPs), or one or more field-programmable gate arrays (FPGAs).

[0110] Optionally, the first processor 2001 can perform various functions of the mobile robot path planning device 410 by running or executing software programs stored in the memory 2002 and calling data stored in the memory 2002.

[0111] In a specific implementation, as one example, the first processor 2001 may include one or more CPUs, for example... Figure 4 CPU0 and CPU1 are shown in the diagram.

[0112] In a specific implementation, as one example, the mobile robot path planning device 410 may also include multiple processors, for example... Figure 4 The first processor 2001 and the second processor 2004 are shown in the diagram. Each of these processors can be a single-core processor or a multi-core processor. Here, a processor can refer to one or more devices, circuits, and / or processing cores used to process data (such as computer program instructions).

[0113] The memory 2002 is used to store the software program that executes the present invention, and is controlled by the first processor 2001 to execute it. The specific implementation method can be referred to the above method embodiment, and will not be repeated here.

[0114] Optionally, the memory 2002 may be a read-only memory (ROM) or other type of static storage device capable of storing static information and instructions, random access memory (RAM) or other type of dynamic storage device capable of storing information and instructions, or electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto. The memory 2002 may be integrated with the first processor 2001 or may exist independently and be accessed through the interface circuit of the mobile robot path planning device 410. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0115] The transceiver 2003 is used to communicate with network devices or with terminal devices.

[0116] Alternatively, transceiver 2003 may include a receiver and a transmitter. Figure 4 (Not shown separately). The receiver is used to implement the receiving function, and the transmitter is used to implement the transmitting function.

[0117] Optionally, the transceiver 2003 can be integrated with the first processor 2001, or it can exist independently and be connected to the interface circuit of the mobile robot path planning device 410. Figure 4 (Not shown in the image) is coupled to the first processor 2001, and this embodiment of the invention does not specifically limit this.

[0118] It should be noted that, Figure 4 The structure of the mobile robot path planning device 410 shown does not constitute a limitation on the router. Actual knowledge structure recognition devices may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0119] Furthermore, the technical effects of the mobile robot path planning device 410 can be referenced from the technical effects of the mobile robot path planning method that integrates wheel-ground interaction characteristics as described in the above method embodiments, and will not be repeated here.

[0120] It should be understood that the first processor 2001 in the embodiments of the present invention may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0121] It should also be understood that the memory in the embodiments of the present invention can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0122] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. A semiconductor medium can be a solid-state drive.

[0123] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0124] In this invention, "at least one" means one or more, and "more than one" means two or more. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of a single item or a plurality of items. For example, at least one of a, b, or c can represent: a, b, c, ab, ac, bc, or abc, where a, b, and c can be a single item or multiple items.

[0125] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0126] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0127] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0128] In the several embodiments provided by this invention, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0129] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0130] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0131] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0132] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A path planning method for a mobile robot that integrates wheel-ground interaction characteristics, characterized in that, The method includes: S1. Initialize the offline pre-computed global geographic environment feature map, the pre-trained traffic energy consumption predictor, and the 3D pose of the robot's starting point and target point; S2. Based on the robot's current 3D pose, generate a set of discrete motion primitives that conform to the robot's nonholonomic constraints using a predefined Ackerman steering geometry model, and calculate the position, heading, and intended action of all feasible candidate child nodes in the next moment based on the discrete motion primitives. S3. Based on the global geographic environment feature map, for each candidate child node and its corresponding intended action, a six-dimensional spatiotemporal feature input vector is constructed. The six-dimensional spatiotemporal feature input vector is then input into a pre-trained traffic energy consumption predictor to obtain the predicted traffic energy consumption. S4. Optimize the action safety constraints of the action to be executed to obtain the action safety penalty term. Based on the three-dimensional travel distance from the robot to the candidate sub-node, the predicted travel energy consumption, the geometric feature resistance, and the action safety penalty term, obtain the single-step energy consumption cost of each candidate sub-node. S5. Select the optimal child node based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm, generate the trajectory sequence, and complete the robot path planning.

2. The mobile robot path planning method integrating wheel-ground interaction characteristics according to claim 1, characterized in that, The six-dimensional spatiotemporal feature input vector in S3 includes: four-dimensional historical dynamic features and two-dimensional forward geometric features; The four-dimensional historical dynamic features include: the mean pitch angle, mean roll angle, standard deviation of pitch angle, and standard deviation of roll angle within a preset time window of the robot body, which are used to implicitly characterize the physical and mechanical properties of the soil. The two-dimensional forward geometric features include: the local roughness and local height difference of candidate child nodes read from the global geographic environment feature map, which are used to explicitly characterize the terrain morphology of the candidate child node location.

3. The mobile robot path planning method integrating wheel-ground interaction characteristics according to claim 1, characterized in that, The pre-training process of the traffic energy consumption prediction machine in S1 includes: A robot dynamics model is established in a multibody dynamics simulation engine, and a soft soil contact model is introduced to generate unstructured terrain through a randomized parameter set. By controlling the robot to perform diverse discrete motion primitives in the unstructured terrain, the robot's real physical response and energy consumption are obtained, and a high-confidence physical dataset of state-action-energy consumption is constructed. Based on a high-confidence physical dataset of state-action-energy consumption, a deep neural network model based on bidirectional long short-term memory network and attention mechanism is trained to obtain a pre-trained travel energy consumption predictor.

4. The mobile robot path planning method integrating wheel-ground interaction characteristics according to claim 3, characterized in that, The training of a deep neural network model based on a bidirectional long short-term memory network and an attention mechanism using a high-confidence physical dataset of state-action-energy consumption includes: Based on the high-confidence physics dataset of state-action-energy consumption, four-dimensional historical dynamic features and two-dimensional forward geometric features are obtained; The four-dimensional historical dynamic features are input into the first bidirectional long short-term memory network to obtain the forward hidden state and the backward hidden state. The forward hidden state and the backward hidden state are concatenated to obtain the soil feature encoding sequence. The two-dimensional forward geometric features are used as the query vector, and the soil feature encoding sequence is used as the key vector and value vector. The context vector containing physical perception is generated by weighted summation. The four-dimensional historical dynamic features are input into the second bidirectional long short-term memory network to obtain the temporal hidden feature vector. The temporal hidden feature vector is used as the query vector, the context vector and the two-dimensional forward geometric features are concatenated as the key vector, the proposed action corresponding to the candidate child node is used as the value vector, and the predicted passage energy consumption is output through the fully connected layer.

5. The mobile robot path planning method integrating wheel-ground interaction characteristics according to claim 1, characterized in that, The action safety constraint optimization for the action to be executed in S4 includes: Based on four-dimensional historical dynamics features, it identifies whether the action to be performed belongs to high-risk motion behavior. For the identified high-risk motion behavior, a nonlinear high penalty factor is applied or adaptive pruning is triggered to remove the high-risk motion behavior.

6. The mobile robot path planning method integrating wheel-ground interaction characteristics according to claim 1, characterized in that, The step S5, which selects the optimal child node based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm to generate a trajectory sequence, includes: A global topological potential field graph is generated using a reverse search algorithm. Based on the global topological potential field graph, the Euclidean distance and topological potential field distance from the candidate child nodes to the target point are obtained. The maximum value of the Euclidean distance and the topological potential field distance is selected as the heuristic prediction cost. The optimal child node is selected from the candidate child nodes based on the sum of the single-step energy consumption cost and the heuristic prediction cost, and a trajectory sequence is generated.

7. A mobile robot path planning device integrating wheel-ground interaction characteristics, wherein the mobile robot path planning device integrating wheel-ground interaction characteristics is used to implement the mobile robot path planning method integrating wheel-ground interaction characteristics as described in any one of claims 1-6, characterized in that, The device includes: The initialization module is used to initialize the offline pre-computed global geographic environment feature map, the pre-trained traffic energy consumption predictor, and the 3D pose of the robot's starting point and target point; The candidate child node calculation module is used to generate a set of discrete motion primitives that conform to the nonholonomic constraints of the robot based on the robot's current three-dimensional pose using a predefined Ackerman steering geometry model, and to calculate the position, heading and intended action of all feasible candidate child nodes in the next moment based on the discrete motion primitives. The traffic energy consumption prediction module is used to construct a six-dimensional spatiotemporal feature input vector for each candidate child node and its corresponding intended action based on the global geographic environment feature map. The six-dimensional spatiotemporal feature input vector is then input into a pre-trained traffic energy consumption prediction machine to obtain the predicted traffic energy consumption. The single-step energy consumption cost calculation module is used to optimize the action safety constraints of the action to be executed and obtain the action safety penalty term. Based on the three-dimensional travel distance from the robot to the candidate sub-node, the predicted travel energy consumption, the geometric feature resistance, and the action safety penalty term, the single-step energy consumption cost of each candidate sub-node is obtained. The output module is used to select the optimal child node based on the single-step energy consumption cost of each candidate child node and the reverse search algorithm, generate a trajectory sequence, and complete the robot path planning.

8. The mobile robot path planning device integrating wheel-ground interaction characteristics according to claim 7, characterized in that, The six-dimensional spatiotemporal feature input vector includes: four-dimensional historical dynamic features and two-dimensional forward geometric features; The four-dimensional historical dynamic features include: the mean pitch angle, mean roll angle, standard deviation of pitch angle, and standard deviation of roll angle within a preset time window of the robot body, which are used to implicitly characterize the physical and mechanical properties of the soil. The two-dimensional forward geometric features include: the local roughness and local height difference of candidate child nodes read from the global geographic environment feature map, which are used to explicitly characterize the terrain morphology of the candidate child node location.

9. A mobile robot path planning device, characterized in that, The mobile robot path planning device includes: processor; A memory storing computer-readable instructions that, when executed by the processor, implement the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1 to 6.

Images