A method for generating a stable trajectory of a vehicle model robot under a graph domain representation network

Abstract

The application discloses a vehicle robot stable trajectory generation method under a graph domain representation network, and comprises the following steps: modeling an environment where a vehicle robot is located as a grid map and converting the grid map into a Euclidean signed distance field, encoding a starting state and an ending state of a trajectory, and splicing to obtain a splicing tensor; generating an original path by using a trained graph domain representation network, wherein in the graph domain representation network, network input is converted into a high-dimensional hidden feature vector by a feature extraction layer, the high-dimensional hidden feature vector is subjected to a global distribution layer to obtain a probability distribution diagram of a path state point in an environment, and a local correction layer corrects the path point according to the high-dimensional hidden feature vector and the probability distribution diagram to obtain the original path; applying double-layer polynomial representation to the original path to initialize trajectory parameters, and obtaining an initial trajectory in a continuous space; taking the initial trajectory as an initial value to perform space-time optimization, and applying an L-BFGS solver to perform fast gradient descent to obtain an executable trajectory.

Description

Technical Field

[0001] This invention belongs to the field of autonomous navigation technology for vehicle-type robots, and particularly relates to a method for generating stable trajectories for vehicle-type robots under a graph domain representation network. Background Technology

[0002] With the continuous advancement of airborne computing systems, vehicle robots have been integrated into various fields of human production activities and public life. In the field of autonomous vehicle systems, motion planning has become a key research direction, attracting widespread attention from academia and industry. Ideal motion planning not only ensures the planning of a safe trajectory connecting the expected start and end points in complex environments, but its stability is also a crucial consideration. This includes two aspects: 1. Stability of computation time: Researchers expect computation time to remain as stable as possible and at a low level regardless of environmental complexity. Good temporal stability enhances the system's adaptability in complex environments, reduces the uncertainty and risk of navigation systems, and thus improves the reliability of autonomous systems. For example, in autonomous vehicles, if motion planning cannot be completed within the expected time, it may lead to a delay in the response of the vehicle's decision-making module, thereby increasing the risk of accidents. Furthermore, temporal stability makes the planner's behavior more predictable, facilitating accurate evaluation of the robot's actions, which is crucial for collaboration between different modules within the system or for coordinating multi-robot tasks. 2. Stability of trajectory solution: Regardless of the complexity of the scenario or the difficulty of the planning problem, the planner should consistently converge to a high-quality feasible solution. This is fundamental to ensuring that the robot navigates to its destination quickly, stably, and safely. In addition to obstacle avoidance requirements, the trajectory should also be kinematically feasible and smooth to minimize energy consumption and mechanical wear on the robot platform, while reducing the trajectory tracking burden on downstream controllers.

[0003] Since direct numerical optimization of trajectories in the continuous domain often leads to infeasible local optima, trajectory planning often adopts a hierarchical framework. That is, firstly, a guiding path is obtained in a lower-dimensional space (i.e., front-end: path planning). On this basis, numerical optimization is then applied to optimize this relatively coarse path into a high-quality, smooth trajectory that satisfies the robot's higher-order kinematic constraints (i.e., back-end: trajectory optimization).

[0004] Traditional front-end planning is typically modeled as a combinatorial optimization problem, with algorithms solving paths by searching or sampling in a low-dimensional configuration space. However, these methods perform poorly in complex environments, easily generating a large number of invalid samples and consuming excessive computational and storage resources. Furthermore, to ensure the quality of the initial path and its completeness in narrow environments, it is often necessary to increase the state space resolution, leading to the combinatorial explosion problem, affecting temporal stability and reducing real-time performance.

[0005] In backend trajectory optimization, existing methods often rely excessively on pre-defined rules, limiting their adaptability in complex and variable scenarios. Due to the complexity of nonholonomic kinematics, most strategies require significant simplification of the motion model or discretization of the motion process to maintain algorithm efficiency. However, to ensure high executability and success rate in dense environments, these strategies require meticulous discretization, negatively impacting the time efficiency of the planning process. Recently, fully differentiable trajectory optimization methods in flat spaces have attracted widespread attention due to their efficiency. However, for vehicle robots with nonholonomic kinematic constraints, the method suffers from singularity issues when the velocity is zero, leading to numerical instability and difficulty in converging to a feasible solution. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention discloses a method for generating stable trajectories of vehicle-type robots under graph domain representation networks, which has excellent time stability and robust convergence.

[0007] According to a first aspect of the embodiments of this application, a method for generating stable trajectories of a vehicle-type robot under a graph domain representation network is provided, comprising: The environment in which the vehicle robot is located is modeled as a grid map and converted into a Euclidean symbolic distance field. The starting state and ending state of the trajectory are encoded to obtain a gridded environment map and a spliced ​​tensor of the starting state encoding and the ending state encoding. Using the spliced ​​tensor as input, the original path is generated using a trained graph domain representation network. In the graph domain representation network, the network input is transformed into a high-dimensional latent feature vector through a feature extraction layer. The high-dimensional latent feature vector is then processed by a global distribution layer to obtain a probability distribution map of the path state points in the environment. Based on the high-dimensional latent feature vector and the probability distribution map, a local correction layer locally corrects the path points to obtain the original path. The original path is initialized with trajectory parameters by applying a two-level polynomial representation to obtain the initial trajectory in continuous space; The initial trajectory is used as the initial value for spatiotemporal optimization, and the L-BFGS solver is applied for fast gradient descent to obtain an executable trajectory.

[0008] Furthermore, the splicing tensor is constructed in the following manner: Set the first tensor, set the mask representing the starting point to -1, the mask representing the target point to 1, and set the rest of the positions to 0; The second tensor is used to represent the cosine value of the direction of the vehicle robot at the starting point and the target point, and the third tensor is used to represent the sine value of the direction of the vehicle robot at the starting point and the target point. The rasterized environment map and the first, second, and third tensors are stitched together to obtain the stitched tensor.

[0009] Furthermore, in the graph domain representation network: The feature extraction layer encodes the concatenated tensor into a higher-dimensional latent space through a fully convolutional network to obtain high-dimensional features, and downsamples the size of the high-dimensional features. The high-dimensional features in the latent space are then fused through a transformer encoder. The global distribution layer generates a probability distribution map of the region where each point on the path is located based on the fused features; The local correction layer generates a position offset tensor based on the fused features and probability distribution map, and then performs local correction on each path point. The position offset tensor assigns a position offset and orientation angle relative to an anchor point to each region, and the anchor point is the center point of each region.

[0010] Furthermore, the loss function during training of the graph domain representation network includes: Anchor point classification loss is used to measure the difference between the probability distribution map and the true distribution; Path supervision loss is used to ensure that the planned original path and the true trajectory remain consistent in the state space. Smoothness loss is used to measure the difference in smoothness between the planned original path and the true trajectory in terms of path length and total change in turning angle. Curvature constraint loss is used to ensure that the original planned path satisfies the maximum curvature constraint. Uniform loss is used to measure the uniformity of the distribution of path points on the original planned path. Obstacle avoidance loss is used to ensure that the distance between each path point in the planned original path and obstacles in the environment is greater than a predetermined minimum safe distance.

[0011] Furthermore, before initializing the trajectory parameters by applying a two-layer polynomial representation to the original path, the original path is filtered using a sliding window.

[0012] Furthermore, the trajectory parameters are initialized by applying a two-level polynomial representation to the filtered original path, thus initializing the trajectory... Represented as a by part A curve composed of polynomials of degree n, the curve being determined by the pseudo-arc length corresponding to each segment of the trajectory. sum coefficient matrix Common parameterization, pseudo-arc length It is represented as a one-dimensional and time-uniform A segment polynomial, consisting of each time interval sum coefficient matrix Parameterization This indicates the dimension of the control input, thereby obtaining the initial trajectory in continuous space, wherein the initial trajectory is: , , in, Represents natural basis functions.

[0013] Furthermore, when using the initial trajectory as the initial value for spatiotemporal optimization, the optimization objective is to minimize the trajectory energy and at least satisfy the boundary conditions of the initial and final state constraints of the trajectory, the continuity constraint at the piecewise polynomial connection, the minimum pseudo-velocity constraint, the dynamic feasibility constraint, and the obstacle avoidance constraint.

[0014] According to a second aspect of the embodiments of this application, a device for generating stable trajectories of a vehicle-type robot under a graph domain representation network is provided, comprising: The encoding module is used to model the environment in which the vehicle robot is located as a grid map and convert it into a Euclidean symbolic distance field. It encodes the starting state and ending state of the trajectory to obtain a gridded environment map and a spliced ​​tensor of the starting state encoding and the ending state encoding. The original path generation module is used to generate an original path using the concatenated tensor as input and a trained graph domain representation network. In the graph domain representation network, the network input is transformed into a high-dimensional latent feature vector through a feature extraction layer. The high-dimensional latent feature vector is then processed by a global distribution layer to obtain a probability distribution map of the path state points in the environment. Based on the high-dimensional latent feature vector and the probability distribution map, a local correction layer locally corrects the path points to obtain the original path. The initial trajectory generation module is used to initialize the trajectory parameters of the original path by applying a two-level polynomial representation to obtain the initial trajectory in continuous space. The trajectory optimization module is used to perform spatiotemporal optimization using the initial trajectory as the initial value, and to apply the L-BFGS solver for fast gradient descent to obtain an executable trajectory.

[0015] According to a third aspect of the embodiments of this application, an electronic device is provided, comprising: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the method as described in the first aspect.

[0016] According to a fourth aspect of the embodiments of this application, a computer-readable storage medium is provided that stores computer instructions thereon, which, when executed by a processor, implement the steps of the method as described in the first aspect.

[0017] The technical solutions provided by the embodiments of this application may include the following beneficial effects: (1) This application performs path planning under the graph domain representation network. By introducing a global distribution layer, the initial path and environment are represented in the same graph domain, which is more conducive to learning obstacle avoidance information. Compared with traditional search / sampling methods, it decouples computational complexity and environmental complexity, and enjoys excellent time stability. No matter how complex the environment is, its computation time tends to be constant.

[0018] (2) This application adopts a trajectory representation based on a two-level polynomial in a flat space. While inheriting the improved solution efficiency brought by the property of differential flatness, it performs smooth remapping on the differential flatness model. This not only efficiently generates high-quality collision-free trajectories, but also fundamentally solves the inherent singularity problem of differential flatness, thereby improving the solution stability and quality of numerical optimization. This application can directly output a reasonable path topology from the environment in milliseconds without the need for explicit sampling or search processes.

[0019] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0021] Figure 1 This is a schematic diagram illustrating a method for generating stable trajectories of a vehicle-type robot under a graph domain representation network, according to an exemplary embodiment.

[0022] Figure 2 This is a schematic diagram of path planning in a simulation environment.

[0023] Figure 3 This is a schematic diagram of trajectory optimization in a simulation environment.

[0024] Figure 4 This is a block diagram illustrating a stable trajectory generation device for a vehicle-type robot under a graph domain representation network, according to an exemplary embodiment.

[0025] Figure 5 This is a schematic diagram of an electronic device according to an exemplary embodiment. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0027] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0028] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0029] Figure 1 This is a schematic diagram illustrating the overall principle and steps of the present invention. The present invention proposes a method for generating stable trajectories for vehicle-type robots using graph domain representation networks. This method is a learning-enhanced approach that benefits from both temporal and solution stability in vehicle-type robot motion planning, and may include the following steps: Step S1: Model the environment in which the vehicle robot is located as a grid map and convert it into a Euclidean symbolic distance field. Encode the starting state and ending state of the trajectory to obtain a gridded environment map and a spliced ​​tensor of the starting state encoding and the ending state encoding. The environment in which the vehicle-type robot operates is modeled as a global grid map, and further calculated as an Euclidean Signed Distance Function (ESDF). Finally, this is converted into a fixed-size tensor form required by the network. The start and end states are also encoded as tensors of the same size as the environment tensor, and concatenated with these tensors to obtain the input required by the network.

[0030] Specifically, the environment Represented by ESDF, where each element represents the symbolic distance from the location to the obstacle. This method employs a novel start-target encoding strategy, which uses a size of This is achieved by applying a predefined first tensor to the corresponding positions on the tensor, thus representing the entire path planning problem on the graph domain. Here, H and W are related to the size of the global map, and their values ​​are the map's width and height divided by the map resolution, respectively. Specifically, the mask representing the starting point is set to -1, the mask representing the target point is set to 1, and the remaining positions in the tensor are set to 0. To fully represent... The space introduced two additional Tensors are used to capture the cosine values ​​of the robot's directions at the starting and target points, respectively; one is used to capture the sine values ​​of the robot's directions at the starting and target points. These are the second and third tensors. The environment described above is then represented. and the beginning and the goal The representations of states are connected to form a The concatenated tensor is then used as input to the graph domain representation network.

[0031] Step S2: Using the spliced ​​tensor as input, the original path is generated using the trained graph domain representation network. In the graph domain representation network, the network input is transformed into a high-dimensional latent feature vector through the feature extraction layer. The high-dimensional latent feature vector is then processed by the global distribution layer to obtain the probability distribution map of the path state points in the environment. Based on the high-dimensional latent feature vector and the probability distribution map, the local correction layer performs local correction on the path state points to obtain the original path. Unlike existing methods that continuously explore and maintain a growth tree from an initial state, this method is more intelligent and direct. Its core idea draws inspiration from the human ability to quickly identify feasible paths from a top-down perspective—once global environmental information is acquired, humans do not need to explore from a first-person perspective at the starting point, but instead directly plan a smooth curve connecting the start and end points. Based on this mechanism, we use neural networks to simulate this cognitive process, achieving efficient path generation through macro-environmental information and expert knowledge, without any sampling or search.

[0032] Specifically, the graph domain representation network encodes the starting state. End-point status coding and rasterized environment maps (size is) As input, it directly outputs a path serialized from multiple state points. ,in Let i be a path state point, 1 ≤ i ≤ N, where N is the number of state points in the path. The resolution of each grid cell is defined as... Furthermore, each state point in the path is associated with the robot's... The states are associated with each other, including their position and heading angle. In terms of network structure, the graph representation network comprises three main components: a feature extraction layer, a global distribution layer, and a local correction layer. In practice, directly locating the position of a specific state point in the environment is challenging and laborious. Therefore, this method employs a two-stage inference architecture.

[0033] First, the environment is extracted through a feature extraction layer. Evenly divided into There are 10 regions, each corresponding to a region in the environment. The system divides the data into blocks of varying sizes. The center point of each region is designated as an anchor point, and the number of regions is a manually determined hyperparameter. Subsequently, a global distribution layer is used to obtain the probability distribution of each state point relative to its respective region. This is represented as a probability graph, thus unifying the representation structure with the environment. Next, a local correction layer is used to further obtain the precise location of the state points based on environmental features and the previously obtained probability distribution. Intuitively, the goal of this two-stage inference architecture is to first roughly determine the region where each state point is located, and then regress its positional offset relative to the anchor point, thereby recovering the global location of the point.

[0034] The feature extraction layer consists of a Fully Convolutional Network (FCN) and a Transformer encoder. The FCN encodes the concatenated tensor into a high-dimensional latent space and downsamples the size of this latent feature to... To match the shape of the region probability distribution This represents the human-defined latent feature dimension. The Transformer encoder further fuses features within the latent space without altering the shape of the feature tensor, thus obtaining the output of the feature extraction layer. This data is then further fed into the global distribution layer to generate a probability distribution map of the regions where each point on the path is located. Its size is Here, this method will... The point belongs to the th point The probability of each region is expressed as: Probability distribution diagram and potential characteristics After connection, the input is fed into the local correction layer, which outputs a value of [size missing]. tensor This tensor The physical meaning is to assign a position offset relative to the anchor point to each region. and direction angle Intuitively, this method uses probability distribution plots... Select the region with the highest probability and offset it from the position tensor. The corresponding position offset and orientation angle are retrieved to accurately reconstruct each point on the path. state.

[0035] Next, this method provides implementation details for each part of the network: (1) Feature Extraction Layer: The FCN consists of four double convolutional layers. The first double convolutional layer does not change the width and height of the tensor, only increasing the number of channels. The subsequent three double convolutional layers downsample the width and height of the tensor and double the number of channels after each layer. Furthermore, MaxPool is used for downsampling, and ReLU and batch normalization are nested between convolutions to enhance generalization performance. The output of the FCN is then redistributed row by row. The size of the data is calculated and, after being positionally encoded, is fed into the Transformer encoder, which learns the relationships between latent features through multi-head self-attention and multilayer perceptron blocks.

[0036] (2) Global Distribution Layer: This module predicts the probability distribution of N state points in the downsampling space and obtains the original path from it. Therefore, the output tensor dimension should be [missing information]. Therefore, this method uses two layers. This operation is achieved efficiently using padded convolutions. Furthermore, this method performs Softmax on each channel of the output tensor to ensure correct probability normalization.

[0037] (3) Local correction layer: Similar to the previous module, this method uses two layers. Convolution assigns a position offset and orientation angle relative to the anchor point to each region, outputting a position offset tensor with dimension . Furthermore, this method applies the Sigmoid activation function to limit the positional offset of the network output, ensuring that the corrected points remain near the anchor points. After obtaining the positional offset tensor, simply adding it to the original path output by the global distribution layer yields the final corrected path.

[0038] Specifically, this invention aims to plan a path that satisfies nonholonomic kinematic constraints and basic obstacle avoidance requirements, while also learning the underlying behavioral strategy of the ground truth trajectory. To achieve this goal, this invention defines the following loss function used in the training process. and with These represent the weights corresponding to each loss function. The network loss functions in this invention include the following: (a) Anchor point classification loss ): This method models the effect of the global distribution layer as a multi-class classification problem. The goal is to maximize the predicted probability of regions containing ground truth trajectory points. To this end, this method uses cross-entropy loss to measure the difference between the probability distribution of predicted regions and the ground truth distribution.

[0039] in, This represents the true probability distribution.

[0040] (b) Path supervision loss ): To ensure that the planned path and the true trajectory remain consistent in the state space, this method uses mean squared error (MSE) loss to supervise each sampling point on the path. (2) Posture:

[0041] in, Represents the corresponding sampling point on the true value trajectory (2) Pose. Considering the periodicity of the angle, this method supervises the sine and cosine values ​​of the angle separately, rather than directly supervising the angle itself.

[0042] (c) Smoothness loss ): Path smoothness is an important indicator for evaluating path quality. This method measures smoothness from two aspects: path length and total turning angle variation. Considering the number of sampling points... When the angle is sufficiently large, the state difference between adjacent sampling points is small. This method uses the difference between the sine and cosine values ​​of adjacent points to approximate the angle difference, thus avoiding problems caused by angle periodicity. To focus on optimizing cases where the smoothness of the planned path is insufficient relative to the true trajectory, this method designs the following loss function to penalize behaviors that deviate from the smoothness of the true trajectory:

[0043]

[0044]

[0045]

[0046] in, and These represent the distances between adjacent sampling points on the planned path and the true trajectory, respectively. (2) State differences.

[0047] (d) Curvature constraint loss ): To ensure that the planned path meets the maximum curvature To address the constraints, this method defines the following loss function:

[0048] The loss function penalizes the portion that exceeds the maximum curvature constraint.

[0049] (e) Uniform loss ): In terms of spatial distribution, this method aims for the sampling points along the planned path to be distributed as evenly as possible. To this end, this method introduces variance to measure the uniformity of the sampling point distribution and normalizes the variance using the path length of the ground truth trajectory to avoid the network becoming overly biased towards long-distance path planning problems.

[0050] (f) Obstacle avoidance loss ): To ensure the safety of the planned path and avoid collisions with obstacles, this method designs the following loss function:

[0051] in, Indicates consideration of robot shape In this case, the signed distance from the state point to the obstacle. This is the minimum safe distance set by the user. The signed distance at any grid point can be directly obtained from the pre-calculated Euclidean signed distance field (ESDF), and bilinear interpolation is used to smooth the discontinuities caused by spatial discretization.

[0052] Through the combined effect of the above loss functions, the network can learn a planned path that satisfies various constraints and approximates the true trajectory, such as... Figure 2 As shown. Figure 2 Trajectory ① is the path output by the traditional dynamic search, trajectory ② is the truth path, and trajectory ③ is the path output by this method. It can be seen that the path output by this method is closer to the truth path and has higher quality.

[0053] To reduce outliers and increase path smoothness, the original path can be filtered using a sliding window in the SE(2) space (such as mean filtering).

[0054] Step S3: Initialize the trajectory parameters by applying a two-level polynomial representation to the original path to obtain the initial trajectory in continuous space; Specifically, the vehicle robot is a differentially flat system, meaning its system state can be analytically represented by a set of flat outputs and their finite-dimensional derivatives. The center position of the rear wheel is selected. As a flat output, the other higher-order states during the robot's motion can be represented as: , , , , in, and These represent the first, second, and third derivatives of the flat output with respect to time, respectively. It is an auxiliary antisymmetric matrix used to simplify calculations. Furthermore, This represents the longitudinal velocity relative to the vehicle's coordinate system. Indicates longitudinal acceleration. Indicates the front wheel steering angle. Indicates the steering angular velocity. This represents the wheelbase. While the differential flatness property can accelerate the optimization process, it's worth noting that when the velocity approaches zero, certain states (e.g.) and Singularities may occur. To fundamentally eliminate the singularity problem, a pseudo-arc length is introduced to indirectly derive the flat model: , in, Let be the pseudo-arc length. Based on the pseudo-arc length-time mapping relationship, the finite-order derivative of the flat output can also be derived as follows:

[0055]

[0056]

[0057] in, and These represent the first, second, and third derivatives of the flat output with respect to the pseudo-arc length, respectively. It is important to note that... This represents the robot's actual speed of movement, while It is defined as a pseudo-velocity. Furthermore, The physical meaning is the magnitude of the velocity along the pseudo-arc length. Based on the above definition, the flat model is modified as follows: , .

[0058] When the robot's speed is zero, this method can Set to zero while maintaining pseudo-velocity The denominator is non-zero. Therefore, the denominator in the formula is strictly positive, thus eliminating the original singularity. Next, this method will describe in detail the method based on the two-level piecewise polynomial. The parameterized form.

[0059] To ensure that the trajectory has sufficient degrees of freedom and smoothness, this method will... and time mapping The parameterization is performed in piecewise polynomial form, thus constructing a compact nonlinear optimization problem. This method transforms the trajectory... Represented as a by part A curve composed of polynomials of degree n, the curve being determined by the pseudo-arc length corresponding to each segment of the trajectory. sum coefficient matrix Common parameterization. Among them, This represents the dimension of the control input. In practice, the value of M can be determined based on the trajectory length and actual requirements. Similarly, pseudo-arc length... It is represented as a one-dimensional and time-uniform A segment polynomial, consisting of each time interval sum coefficient matrix Parameterization. Furthermore, this method considers the strict correspondence between each trajectory segment in the two-layer piecewise polynomial, i.e., for each time interval... The robot should move the corresponding pseudo-arc length:

[0060] Based on the above modeling, the initial trajectory The A segment can be represented as: , , in, This represents the natural basis functions.

[0061] Step S4: Use the initial trajectory as the initial value for spatiotemporal optimization, and apply the L-BFGS solver for fast gradient descent to obtain an executable trajectory; Furthermore, due to the strict correspondence between each segment of the bilayer polynomial, this method can be further derived as follows:

[0062] in, It was set to 0. Furthermore, Segmented polynomial It can be represented as:

[0063] Combining various constraints, and based on the modified flat model with the addition of a first-order time regularization, the original trajectory optimization problem can be formulated as a nonlinear constrained optimization problem that simultaneously minimizes trajectory energy and execution time.

[0064]

[0065]

[0066]

[0067] in, This represents the weight of the user-defined time regularization term, used to constrain the total time. . formula The boundary conditions represent the initial and final state constraints of the trajectory. Formula This represents the continuity constraint at the junction of a piecewise polynomial. Formula This is a minimum pseudo-velocity constraint introduced to avoid singularities, where It is a threshold. This method incorporates various inequality constraints commonly found in trajectory planning problems, including dynamic feasibility constraints and obstacle avoidance constraints. To solve this constrained nonlinear optimization problem, this method first reconstructs the original problem using the minimum energy condition of the trajectory and eliminates its equality constraints without sacrificing optimality. Subsequently, this method employs the Augmented Lagrange Multiplier Method (ALM) to relax the inequality constraints. This method iteratively updates the dual variables while using the efficient finite-memory Broyden–Fletcher–Goldfarb–Shanno (L-BFGS) algorithm to solve the approximate unconstrained problem. The trajectory optimized by this method is visualized as follows: Figure 3 As shown. Figure 3 Trajectory ① in the diagram is the trajectory optimized by this method, while trajectory ② is the trajectory optimized by the traditional flat class method. In comparison, due to the elimination of singular points, the trajectory optimized by this method is smoother.

[0068] In summary, this application represents the original path planning problem in a graph domain and specifically utilizes the Transformer model to directly learn the optimal path from the environment, achieving a constant computation time across various scenarios. Thanks to the unified representation paradigm in the graph domain, this method eliminates the need for learning map scale and improves the understanding of the geometric relationship between the trajectory and the environment, thus enhancing obstacle avoidance performance. For backend optimization, a specially designed two-layer trajectory representation method is introduced, effectively combining differential flatness to achieve efficient solution. Furthermore, this representation method addresses the singularity problem inherent in flat models, thereby enhancing numerical stability and ensuring robust generation of feasible solutions.

[0069] Corresponding to the aforementioned embodiments of the method for generating stable trajectories of vehicle robots under graph domain representation networks, this application also provides embodiments of a device for generating stable trajectories of vehicle robots under graph domain representation networks.

[0070] Figure 4 This is a block diagram illustrating a stable trajectory generation device for a vehicle-type robot under a graph domain representation network, according to an exemplary embodiment. (Refer to...) Figure 4 The device may include: The encoding module 21 is used to model the environment where the vehicle robot is located as a grid map and convert it into a Euclidean symbolic distance field, and to encode the starting state and ending state of the trajectory to obtain a gridded environment map and a splicing tensor of the starting state encoding and the ending state encoding. The original path generation module 22 is used to generate an original path using the spliced ​​tensor as input and a trained graph domain representation network. In the graph domain representation network, the network input is transformed into a high-dimensional latent feature vector through a feature extraction layer. The high-dimensional latent feature vector is then processed by a global distribution layer to obtain a probability distribution map of the path state points in the environment. Based on the high-dimensional latent feature vector and the probability distribution map, a local correction layer locally corrects the path points to obtain the original path. The initial trajectory generation module 23 is used to initialize the trajectory parameters of the original path by applying a two-level polynomial representation to obtain the initial trajectory in continuous space. The trajectory optimization module 24 is used to perform spatiotemporal optimization using the initial trajectory as the initial value, and to apply the L-BFGS solver for fast gradient descent to obtain an executable trajectory.

[0071] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0072] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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 modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0073] Accordingly, this application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the method for generating stable trajectories of a vehicle robot under a graph domain representation network as described above.

[0074] Accordingly, this application also provides an electronic device, including: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method for generating stable trajectories of a vehicle-type robot under a graph domain representation network as described above. Figure 5 The diagram shown is a hardware structure diagram of any device with data processing capabilities, which is the stable trajectory generation device for a vehicle-type robot under a graph domain representation network provided in an embodiment of the present invention. (Except for...) Figure 5 In addition to the processor, memory, and network interface shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0075] Accordingly, this application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the method for generating a stable trajectory of a vehicle-type robot under a graph domain representation network as described above. The computer-readable storage medium can be an internal storage unit of any device with data processing capabilities as described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any device with data processing capabilities and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the device with data processing capabilities, and can also be used to temporarily store data that has been output or will be output.

[0076] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0077] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A method for generating stable trajectories of a vehicle-type robot under a graph domain representation network, characterized in that, include: The environment in which the vehicle robot is located is modeled as a grid map and converted into a Euclidean symbolic distance field. The starting state and ending state of the trajectory are encoded to obtain a gridded environment map and a spliced ​​tensor of the starting state encoding and the ending state encoding. Using the spliced ​​tensor as input, the original path is generated using a trained graph domain representation network. In the graph domain representation network, the network input is transformed into a high-dimensional latent feature vector through a feature extraction layer. The high-dimensional latent feature vector is then processed by a global distribution layer to obtain a probability distribution map of the path state points in the environment. Based on the high-dimensional latent feature vector and the probability distribution map, a local correction layer locally corrects the path points to obtain the original path. The original path is initialized with trajectory parameters by applying a two-level polynomial representation to obtain the initial trajectory in continuous space; The initial trajectory is used as the initial value for spatiotemporal optimization, and the L-BFGS solver is applied for fast gradient descent to obtain an executable trajectory.

2. The method according to claim 1, characterized in that, The splicing tensor is constructed in the following manner: Set the first tensor, set the mask representing the starting point to -1, the mask representing the target point to 1, and set the rest of the positions to 0; The second tensor is used to represent the cosine value of the direction of the vehicle robot at the starting point and the target point, and the third tensor is used to represent the sine value of the direction of the vehicle robot at the starting point and the target point. The rasterized environment map and the first, second, and third tensors are stitched together to obtain the stitched tensor.

3. The method according to claim 1, characterized in that, In the graph domain representation network: The feature extraction layer encodes the concatenated tensor into a higher-dimensional latent space through a fully convolutional network to obtain high-dimensional features, and downsamples the size of the high-dimensional features. The high-dimensional features in the latent space are then fused through a transformer encoder. The global distribution layer generates a probability distribution map of the region where each point on the path is located based on the fused features; The local correction layer generates a position offset tensor based on the fused features and probability distribution map, and then performs local correction on each path point. The position offset tensor assigns a position offset and orientation angle relative to an anchor point to each region, and the anchor point is the center point of each region.

4. The method according to claim 1, characterized in that, The loss function for training the graph domain representation network includes: Anchor point classification loss is used to measure the difference between the probability distribution map and the true distribution; Path supervision loss is used to ensure that the planned original path and the true trajectory remain consistent in the state space. Smoothness loss is used to measure the difference in smoothness between the planned original path and the true trajectory in terms of path length and total change in turning angle. Curvature constraint loss is used to ensure that the original planned path satisfies the maximum curvature constraint. Uniform loss is used to measure the uniformity of the distribution of path points on the original planned path. Obstacle avoidance loss is used to ensure that the distance between each path point in the planned original path and obstacles in the environment is greater than a predetermined minimum safe distance.

5. The method according to claim 1, characterized in that, Before initializing the trajectory parameters by applying a two-level polynomial representation to the original path, the original path is filtered using a sliding window.

6. The method according to claim 1, characterized in that, The filtered original path is initialized with trajectory parameters using a two-level polynomial representation. Represented as a by part A curve composed of polynomials of degree n, the curve being determined by the pseudo-arc length corresponding to each segment of the trajectory. sum coefficient matrix Common parameterization, pseudo-arc length It is represented as a one-dimensional and time-uniform A segment polynomial, consisting of each time interval sum coefficient matrix Parameterization This indicates the dimension of the control input, thereby obtaining the initial trajectory in continuous space, wherein the initial trajectory is: ， ， in, This represents the natural basis functions.

7. The method according to claim 1, characterized in that, When the initial trajectory is used as the initial value for spatiotemporal optimization, the optimization objective is to minimize the trajectory energy and at least satisfy the boundary conditions of the initial and final state constraints of the trajectory, the continuity constraint at the piecewise polynomial connection, the minimum pseudo-velocity constraint, the dynamic feasibility constraint, and the obstacle avoidance constraint.

8. A device for generating stable trajectories of a vehicle-type robot under a graph domain representation network, characterized in that, include: The encoding module is used to model the environment in which the vehicle robot is located as a grid map and convert it into a Euclidean symbolic distance field. It encodes the starting state and ending state of the trajectory to obtain a gridded environment map and a spliced ​​tensor of the starting state encoding and the ending state encoding. The original path generation module is used to generate an original path using the concatenated tensor as input and a trained graph domain representation network. In the graph domain representation network, the network input is transformed into a high-dimensional latent feature vector through a feature extraction layer. The high-dimensional latent feature vector is then processed by a global distribution layer to obtain a probability distribution map of the path state points in the environment. Based on the high-dimensional latent feature vector and the probability distribution map, a local correction layer locally corrects the path points to obtain the original path. The initial trajectory generation module is used to initialize the trajectory parameters of the original path by applying a two-level polynomial representation to obtain the initial trajectory in continuous space. The trajectory optimization module is used to perform spatiotemporal optimization using the initial trajectory as the initial value, and to apply the L-BFGS solver for fast gradient descent to obtain an executable trajectory.

9. An electronic device, characterized in that, include: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.

Images