Robot trajectory generation method and system based on diffusion and rule inference

Abstract

The application provides a robot trajectory generation method and system based on diffusion and rule inference, and relates to the technical field of autonomous navigation.The application proposes a novel conditional diffusion trajectory generation model based on architecture induction bias, which is not only dependent on loss function for result constraint, but also divides the complex trajectory generation task into two mutually coordinated technical processes:1, based on local environment graph construction, graph convolution propagation and graph level feature convergence, the inference of the overall passable structure and obstacle topological relationship within the scope of the current local environment graph is realized;2, based on node level scene representation, target feature continuous injection and prediction step by step context update, the fine-grained local maneuvering control of obstacle avoidance, turning and end convergence behavior at each future waypoint is realized, and the high-fidelity generation of the trajectory in the complex unstructured environment is ensured.Meanwhile, a GTGRU is designed, the end drift problem is solved by introducing time attention bias.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation technology, specifically to a method and system for generating robot trajectories based on diffusion and rule inference. Background Technology

[0002] With the rapid development of intelligent logistics and service robots, autonomous mobile robots have been widely applied in various scenarios such as warehousing and handling, outdoor inspection, and home services. Their operating environments are also gradually expanding from structured factory workshops and warehouse corridors to large-scale, unstructured, and complex scenarios, such as urban roads, wilderness environments, and large venues. Unstructured environments are characterized by complex terrain, randomly distributed obstacles, and vast spatial ranges, placing higher demands on the trajectory generation capabilities of autonomous mobile robots compared to structured environments. Generating safe, smooth, and feasible trajectories in large-scale unstructured environments is a prerequisite for the operation of autonomous mobile robots.

[0003] Existing trajectory generation methods in unstructured environments mainly include data-driven trajectory generation methods. These methods typically follow two core paradigms when addressing the trade-off between trajectory performance and control requirements: The first is classical generative models or neural network methods with embedded optimization modules. The core design goal of these methods is to enhance the geometric constraints and end-effector accuracy of the trajectory. However, these methods are prone to kinematic discontinuities in the trajectory, affecting the stability of robot motion. The second is the rapidly developing trajectory generation method based on diffusion models. This type of method has significant advantages in generating globally coherent and kinematically feasible trajectories. Relying on the probabilistic generation characteristics of diffusion models, it can generate trajectories that conform to robot kinematic constraints (such as turning radius and velocity / acceleration limits), and the trajectories are globally smooth and coherent, requiring no additional complex post-processing optimization. It is particularly suitable for weakly constrained navigation scenarios (such as scenarios without a clear fixed endpoint and with frequent dynamic environmental changes).

[0004] However, existing trajectory generation methods based on diffusion models cannot dynamically adjust environmental information, and the randomness of the model itself leads to insufficient end-point control accuracy. They cannot balance trajectory smoothness and end-point accuracy, making it difficult to meet the requirements of high-precision operations. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a robot trajectory generation method and system based on diffusion and rule inference, which solves the technical problem that existing trajectory generation methods based on diffusion models cannot dynamically adjust environmental information.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: In a first aspect, the present invention provides a robot trajectory generation method based on diffusion and rule inference. The robot trajectory is generated through a pre-trained conditional diffusion trajectory generation model based on architecture inductive bias. The conditional diffusion trajectory generation model includes a relational inductive bias network, a temporal inductive bias network, and a denoising network. The robot trajectory generation method includes: Obtain the current LiDAR point cloud information, robot historical velocity sequence, and target coordinates; The LiDAR point cloud information, robot historical velocity sequence, and target coordinates are encoded using a relational inductive bias network to obtain LiDAR point cloud information encoding, velocity features, and target features. Graph convolution propagation is then applied to the LiDAR point cloud information encoding to obtain node-level scene representations. Graph pooling is then applied to the node-level scene representations to obtain graph-level LiDAR aggregated features. Finally, the velocity features and graph-level LiDAR aggregated features are fused to obtain the environmental context vector. Under the joint constraints of global environment context vector, target features, and node-level scene representation, in the prediction time domain, the graph conditional time-aware GRU of the temporal inductive bias network is sequentially expanded into layers 1 to m, and the temporal hidden state of the future trajectory is recursively generated along the prediction time domain. , … ; Temporally hidden states are processed using a denoising network. , … The noisy input trajectory is processed to obtain the robot trajectory.

[0007] Preferably, the relation induction bias network includes a structured scene perception module SSP based on GNN and a fused multilayer perceptron; The structured scene perception module includes an encoding layer, a convolutional layer, and a pooling layer; The encoding layer performs raw conditional encoding on the current LiDAR point cloud information, the robot's historical velocity sequence, and the target coordinates, respectively, to obtain the LiDAR point cloud information encoding and velocity features. and target features ; Convolutional layers perform graph convolution propagation on the LiDAR point cloud information encoding to obtain node-level scene representations. ; Pooling layer Graph-level pooling is performed to obtain graph-level LiDAR aggregated features. ; Fusion of speed features from multilayer perceptrons Features aggregated with graph-level LiDAR The elements are fused to form an environment context vector. .

[0008] Preferably, under the joint constraints of the global environment context vector, target features, and node-level scene representation, in the prediction time domain, the graph conditional time-aware GRU of the temporal inductive bias network is sequentially expanded into layers one through m, and the temporal hidden state of the future trajectory is recursively generated along the prediction time domain. , … ,include: Node-level scene representation Together with the current recursive state, they form the context information of the current prediction step. For the first prediction step, based on the initial hidden state... Node-level scene representation Forming the first prediction step context information For the m-th prediction step, based on the hidden state of the previous prediction step... Node-level scene representation Forming context information for the current prediction step ; where the initial hidden state Its expression is: in, Represents the environment context vector The mapping weight matrix to the initial hidden state space. Representing target features The mapping weight matrix to the initial hidden state space. Indicates the initial hidden state; For the first future waypoint, input the current step. Initial hidden state First prediction step context information diffusion time condition t and target characteristics The first-layer graph conditional time-aware GRU is input together, and the first temporal hidden state is obtained through updating. ; For the m-th future path point, input the current step. Previous hidden state Current prediction step context information diffusion time condition t and target characteristics Input the conditional time-aware GRU of the m-th layer graph together to complete the current hidden state update.

[0009] Preferably, the denoising network includes an encoder, a decoder, and an output layer multilayer perceptron; the output layer multilayer perceptron includes a layer normalization layer, a serial fully connected layer, a cumulative summation layer, and a Bézier curve smoothing layer. Among them, the encoder outputs the noisy trajectory of the input. Feature encoding is performed, and the decoder receives and processes the encoding result and the hidden state sequence. , … The trajectory hiding representation is obtained. ; Layer normalized layer for trajectory hiding representation After processing, a stable hidden representation is obtained, denoted as . ; Serial fully connected layers for stable hidden representation Perform sequence mapping to obtain the trajectory increment sequence of all prediction steps; After obtaining the trajectory increments of all prediction steps, the future trajectory coordinate sequence is calculated through a cumulative summation layer; The Bézier curve smoothing layer smooths the future trajectory coordinate sequence to obtain the robot trajectory.

[0010] Preferably, the loss function of the pre-trained conditional diffusion trajectory generation model based on architecture inductive bias during training includes: in, Indicates the total training loss; This represents the loss during diffusion denoising and reconstruction. Represents the k-th type of auxiliary loss term; Indicates the first One training round, Indicates the index of the auxiliary loss term. Indicates the first Class Auxiliary Loss in the 1st Dynamic weights across training rounds.

[0011] Preferably, the auxiliary loss terms include imitation fidelity loss, kinematic consistency loss, and safety constraint loss; Among them, the loss of imitation fidelity include: in, This represents the future trajectory predicted by the model. This represents the actual trajectory of the expert for the corresponding sample. This represents the Hausdorff distance between the predicted trajectory and the expert's actual trajectory. This represents the dynamic time-warped distance between the predicted trajectory and the expert's actual trajectory. This indicates the endpoint of the predicted trajectory. Indicates the endpoint of the expert's actual trajectory; Loss of kinematic consistency include: in, Indicates the first The velocity vector between adjacent future waypoints in a segment. , Indicates the index of all valid waypoints Find the average. This represents a very small positive number introduced to prevent the denominator from being zero; Safety constraint loss include: For any breakthrough point Safety radius trajectory points Apply soft punishment: .

[0012] Preferably, the pre-trained conditional diffusion trajectory generation model based on architecture inductive bias is trained during the training process. The course will be implemented in two phases, specifically including: when This is the initial imitation stage, aiming to improve... The weights corresponding to the loss terms; when In order to improve the efficiency in the subsequent refining stage. and The weights corresponding to the loss terms; where, This indicates the preset training round threshold for switching from the initial imitation stage to the subsequent refinement stage in the course learning process.

[0013] Secondly, the present invention provides a robot trajectory generation system based on diffusion and rule inference. The robot trajectory generation system generates robot trajectories through a pre-trained conditional diffusion trajectory generation model based on architecture inductive bias. The conditional diffusion trajectory generation model includes a relational inductive bias network, a temporal inductive bias network, and a denoising network. The robot trajectory generation system includes: The data acquisition module is used to acquire the current LiDAR point cloud information, the robot's historical velocity sequence, and the target coordinates; The trajectory generation module is used to perform the following operations: The LiDAR point cloud information, robot historical velocity sequence, and target coordinates are encoded using a relational inductive bias network to obtain LiDAR point cloud information encoding, velocity features, and target features. Graph convolution propagation is then applied to the LiDAR point cloud information encoding to obtain node-level scene representations. Graph pooling is then applied to the node-level scene representations to obtain graph-level LiDAR aggregated features. Finally, the velocity features and graph-level LiDAR aggregated features are fused to obtain the environmental context vector. Under the joint constraints of global environment context vector, target features, and node-level scene representation, in the prediction time domain, the graph conditional time-aware GRU of the temporal inductive bias network is sequentially expanded into layers 1 to m, and the temporal hidden state of the future trajectory is recursively generated along the prediction time domain. , … ; Temporally hidden states are processed using a denoising network. , … The noisy input trajectory is processed to obtain the robot trajectory.

[0014] Thirdly, the present invention provides a computer-readable storage medium storing a computer program for generating robot trajectories based on diffusion and rule inference, wherein the computer program causes a computer to execute the robot trajectory generation method based on diffusion and rule inference as described above.

[0015] Fourthly, the present invention provides an electronic device, comprising: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing a robot trajectory generation method based on diffusion and rule inference as described above.

[0016] (III) Beneficial Effects This invention provides a robot trajectory generation method and system based on diffusion and rule inference. Compared with existing technologies, it has the following advantages: This invention proposes a novel conditional diffusion trajectory generation model based on architectural inductive bias. This framework does not rely solely on a loss function for result constraints, but decomposes the complex trajectory generation task into two complementary technical processes: First, based on local environment graph construction, graph convolutional propagation, and graph-level feature convergence, it achieves reasoning about the overall traversable structures and obstacle topology within the current local environment graph. Second, based on node-level scene representation, continuous injection of target features, and step-by-step context updates, it achieves fine-grained local maneuver control over obstacle avoidance, turning, and terminal convergence behaviors at each future waypoint. This solves the technical problem that existing diffusion-based trajectory generation methods cannot dynamically adjust environmental information, ensuring high-fidelity trajectory generation in complex unstructured environments. Simultaneously, this invention designs a graph conditional time-aware GRU (GTGRU), which, by introducing a temporal attention bias, solves the terminal drift problem caused by the reliance on static conditional encoding in standard diffusion planners. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0018] Figure 1 This is a schematic diagram of the structure of the conditional diffusion trajectory generation model in an embodiment of the present invention; Figure 2 To verify the comparison between the predicted trajectories of different models and the actual trajectories in the experiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention are described clearly and completely. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This application provides a robot trajectory generation method and system based on diffusion and rule inference, which solves the technical problem that existing trajectory generation methods based on diffusion models cannot dynamically adjust environmental information. It utilizes a graph conditional time-aware GRU to dynamically focus on fine-grained constraints, achieving centimeter-level end-effector fidelity.

[0021] The technical solution in this application is to solve the above-mentioned technical problems, and the general idea is as follows: Existing trajectory generation methods based on diffusion models are among the most advanced. While these methods can generate globally smooth, coherent trajectories that conform to kinematic constraints and require no complex post-processing, making them suitable for weakly constrained navigation scenarios, they still have the following drawbacks: Existing diffusion models typically encode the environmental context into a fixed vector all at once before trajectory generation begins. This encoding paradigm results in a lack of dynamic perception during iterative denoising, exhibiting a "blind generation" characteristic and failing to dynamically adjust the level of attention to local geometric constraints based on the denoising progress. Due to the inherent stochastic nature of diffusion models and the aforementioned static encoding defects, the models are prone to significant accumulated errors and spatial drift when approaching the target point, making it difficult to meet the centimeter-level positioning accuracy required for tasks such as automatic docking. Furthermore, because existing diffusion models lack a deep understanding of the topological relationships in unstructured scenes, the generated trajectories have a higher collision risk in complex obstacle areas.

[0022] To address the aforementioned issues, this invention provides a conditional generation framework for graph-free trajectory generation, effectively balancing the critical trade-off between geometric accuracy and kinematic smoothness. By introducing a temporal attention bias, it resolves the end-point drift problem caused by the reliance on static conditional encoding in standard diffusion planners. Utilizing a sparse cross-attention mechanism, the generator can dynamically "observe" local obstacles and target states at each time step of denoising, thereby achieving centimeter-level end-point pose fidelity.

[0023] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0024] This invention provides a robot trajectory generation method based on diffusion and rule inference. The method generates robot trajectories using a pre-trained conditional diffusion trajectory generation model based on architecture inductive bias. The conditional diffusion trajectory generation model includes a relational inductive bias network, a temporal inductive bias network, and a denoising network. The robot trajectory generation method includes: Obtain the current LiDAR point cloud information, robot historical velocity sequence, and target coordinates; The LiDAR point cloud information, robot historical velocity sequence, and target coordinates are encoded using a relational inductive bias network to obtain LiDAR point cloud information encoding, velocity features, and target features. Graph convolution propagation is performed on the encoded LiDAR point cloud information to obtain node-level scene representations. ;right Graph-level pooling is performed to obtain graph-level LiDAR aggregated features. ; for velocity characteristics Features aggregated with graph-level LiDAR The environment context vector is obtained by fusion. ; Under the joint constraints of global environment context vector, target features, and node-level scene representation, in the prediction time domain, the graph conditional time-aware GRU of the temporal inductive bias network is sequentially expanded into layers 1 to m, and the temporal hidden state of the future trajectory is recursively generated along the prediction time domain. , … ; Temporally hidden states are processed using a denoising network. , … The noisy input trajectory is processed to obtain the robot trajectory.

[0025] This invention proposes a novel conditional diffusion trajectory generation model based on architectural inductive bias. This framework does not rely solely on a loss function for result constraints, but decomposes the complex trajectory generation task into two complementary technical processes: First, based on local environment graph construction, graph convolutional propagation, and graph-level feature convergence, it achieves reasoning about the overall traversable structures and obstacle topology within the current local environment graph. Second, based on node-level scene representation, continuous injection of target features, and step-by-step context updates, it achieves fine-grained local maneuver control over obstacle avoidance, turning, and terminal convergence behaviors at each future waypoint, ensuring high-fidelity trajectory generation in complex unstructured environments. Simultaneously, this invention designs a graph conditional time-aware GRU (GTGRU), which, by introducing a temporal attention bias, solves the terminal drift problem caused by the reliance on static conditional encoding in standard diffusion planners.

[0026] The following is combined with, for example Figure 1 The schematic diagram of the conditional diffusion trajectory generation model based on architectural inductive bias illustrates the robot trajectory generation method in detail: Conditional diffusion trajectory generation models based on architecture inductive bias include: relational inductive bias network, temporal inductive bias network, and denoising network.

[0027] The relational induction bias network comprises a GNN-based structured scene perception module (SSP) and a fusion multilayer perceptron. The structured scene perception module includes an encoder, convolutional layers (Convs), and pooling layers, used to perform raw conditional encoding on the current LiDAR point cloud information, the robot's historical velocity sequence, and the target coordinates, forming feature representations for the environment perception branch, velocity branch, and target branch, respectively. Specifically, the historical velocity sequence, after encoding, forms velocity features. The target coordinates are encoded to form target features. The LiDAR point cloud information, after being encoded, enters the environment perception branch, and first obtains a node-level scene representation through Convs graph convolution propagation. Subsequently, regarding Graph-level pooling is performed to obtain graph-level LiDAR aggregated features. The This information is used to preserve fine-grained spatial structure information and local obstacle geometry in the current local environment graph, and participates in context computation as a node-level scene representation in subsequent temporal recursion. It should be noted that all information here refers to the overall aggregation of all nodes within the current local environment graph, rather than a global model of the entire scene map.

[0028] For the environmental perception branch, the first step is to analyze the local LiDAR point cloud observed at the current moment. P t Voxel downsampling is performed, and a local environment map is constructed based on the nearest neighbor rule. G This processing reduces the computational burden of directly involving the original point cloud in diffusion denoising, while preserving the local obstacle distribution, the shape of passable space, and the relative geometric relationships between nodes. This allows subsequent graph convolutional propagation to perform local relationship reasoning on a structured topology. The local environment graph corresponds to the spatial structure within the current local observation range, rather than a global map of the entire scene, and can be represented as: G =( V , E ), in, V This represents the set of nodes in the local environment graph. E This represents the set of adjacent edges between nodes. Each node in the graph corresponds to a local spatial cell obtained after voxel downsampling, and its initial node features are denoted as . Subsequently, the local environment graph is input into the Convs graph convolution propagation process in the structured scene perception module (SSP) to perform layer-by-layer message propagation and feature updates for each node.

[0029] For the Layered graph convolution, nodes The input hidden features are represented as By combining the relative spatial relationships between the central node and its neighboring nodes, the nodes can be obtained. In the The updated hidden features are expressed as follows: in, Represents a node The set of neighboring nodes, This indicates a feature concatenation operation, item It is used to explicitly encode the relative geometric relationships between the central node and its neighboring nodes, so that the learned representation can effectively capture the distribution of local obstacles and spatial topological details.

[0030] Velocity features are obtained by encoding the robot's historical motion state sequence using velocity branching. And by fusing multilayer perceptron and graph-level LiDAR to aggregate features The elements are fused to form an environment context vector. Its expression is: It is worth noting that although the embodiments of the present invention simultaneously receive three types of raw inputs—LiDAR point cloud information, robot historical velocity sequence, and target coordinates—during the front-end encoding stage, they do not directly incorporate the target features into the environmental context vector. Instead of focusing on the middle, it focuses on the target features. As an independent condition, it is retained. Therefore, the relational inductive bias network outputs a node-level scene representation. and environment context vector On the other hand, retaining target features This achieves a separation of responsibilities among environmental perception, motion state constraints, and target guidance. Structurally, this design decouples responsibilities, with the GNN-based structured scene perception module focusing on environmental navigability (satisfying...). ), while target guidance is dynamically processed by the time generation process.

[0031] Temporal inductive bias network: The temporal inductive bias network is set after the relational inductive bias network to be used in the context vector. Target features and node-level scene representation Under the common constraints, the temporal hidden states of future trajectories are recursively generated along the prediction time domain. The temporal inductive bias network uses Graph-Conditioned Time-Aware GRU (GTGRU) as its recursive core, and unfolds sequentially into GTGRU Layer 1 to GTGRU Layer m in the prediction time domain. It should be noted that... Figure 1 In this context, GTGRU Layer 1 to GTGRU Layer m represent the recursive update process corresponding to the 1st to the mth future waypoints, and are not physically stacked multi-layer GRU networks.

[0032] The key to this embodiment of the invention lies not in changing the basic gating principle of the classic GRU, but in continuously introducing two types of conditional information corresponding to the attached figures in each prediction step: First, node-level scene representation Together with the current recursive state, they form the context information of the current prediction step. It is used to dynamically represent the local obstacle constraints most relevant to the current future waypoint; Secondly, target features It not only participates in conditional fusion during the initial state construction phase, but also continuously injects it during the gating update process of each subsequent prediction step, thereby enabling the target guidance to take effect cyclically throughout the entire prediction time domain, rather than being given only once at the start of the recursion.

[0033] Specifically, the environment context vector With target features The two are combined to form global conditional features And construct the initial hidden state accordingly. Its expression is: in, Represents the environment context vector The mapping weight matrix to the initial hidden state space. Representing target features The mapping weight matrix to the initial hidden state space. This indicates the initial hidden state.

[0034] For the first prediction step, based on the initial hidden state... Node-level scene representation Forming the first prediction step context information The aforementioned The context signal corresponding to the input Graph-Conditioned Time-Aware GRU gating structure in the attached figure is used to characterize the dynamic constraint information related to the distribution of local obstacles in the current prediction step.

[0035] For the m-th prediction step, based on the hidden state of the previous prediction step... Node-level scene representation Forming context information for the current prediction step The aforementioned This allows the model to dynamically adjust its focus on the distribution of local obstacles as the prediction step changes.

[0036] For the first future waypoint, input the current step. Initial hidden state First prediction step context information diffusion time condition t and target characteristics Inputting both into Graph-Conditioned Time-AwareGRU updates the first time-series hidden state. .

[0037] For the m-th future path point, input the current step. Previous hidden state Current prediction step context information diffusion time condition t and target characteristics Input the same recursive unit to complete the update of the current hidden state.

[0038] because and The process participates in gating updates at every prediction step, thus enabling it to, on the one hand, rely on... Dynamically introducing local obstacle constraints in the current step, on the other hand, can ensure that target guidance is continuously injected throughout the entire prediction time domain.

[0039] It should be noted that in this embodiment of the invention, the prediction step and the future waypoint are in one-to-one correspondence: the m-th prediction step corresponds to the recursive generation process of the m-th future waypoint, and the prediction step first updates the temporal hidden state of the corresponding future waypoint. Instead of the final trajectory coordinates, the temporal hidden states are output sequentially from GTGRU Layer 1 to GTGRU Layer m after all prediction steps are completed using the above recursive method. , … This forms a hidden state sequence in the future prediction time domain; the hidden state sequence is then input into a denoising network, which completes trajectory reconstruction and final trajectory output.

[0040] Denoising network: The denoising network comprises an encoder, a decoder, and an output multilayer perceptron (MLP), used to perform trajectory denoising and reconstruction guided by the hidden state sequence output by the temporal inductive bias network. Specifically, among them, This represents the noisy trajectory at the current diffusion time step t, whose initial value is random noise and is continuously updated during the progressive denoising iterations. It serves as the input to the denoising network. (It should be noted that during the training process of the conditional diffusion trajectory generation model based on architecture inductive bias,...) (Obtained by diffusion and noise addition from the real future trajectory). Specifically, the Encoder first processes the input noisy trajectory. Feature encoding is performed, and the Decoder receives the encoding result and the hidden state sequence output by the temporal inductive bias network. , … The trajectory hiding representation is obtained. ,Will As input to the subsequent Output MLP.

[0041] Subsequently, the Output MLP performs normalization, sequence mapping, and trajectory increment prediction sequentially on the trajectory hidden representation. The stable hidden representation after layer normalization (LayerNorm) is denoted as... ,Right now: Building upon this, a sequential fully connected layer performs sequence mapping on the stable hidden representation, outputting the trajectory increment sequence for all prediction steps, where the output of the m-th prediction step is denoted as ( ).

[0042] After obtaining the trajectory increments for all prediction steps, the future trajectory coordinate sequence can be obtained through a cumulative summation operation (cumsum). x , y ), Furthermore, to ensure the generated trajectory meets higher requirements for smoothness and executability, the trajectory coordinate sequence can be smoothed using a Bezier curve to obtain the final robot trajectory. trajectory Its expression is: in, Indicates the first m The coordinates of the future waypoints are then determined. Thus, the temporal inductive bias network is responsible for generating the hidden state sequence in the future prediction time domain, while the denoising network, guided by this hidden state sequence, further maps the hidden states into trajectory increments, trajectory coordinates, and the final smoothed trajectory result through an Encoder, Decoder, and Output MLP.

[0043] The loss function settings during the training process of the pre-trained conditional diffusion trajectory generation model based on architecture inductive bias in this embodiment of the invention are as follows: Unlike the standard diffusion model that predicts noise residuals, this embodiment of the invention parameterizes the generator to directly predict noise-free trajectories. This form allows differentiable geometric and kinematic constraints to be directly applied to the output space. The overall training objective is jointly constituted by the standard diffusion denoising and reconstruction loss and the auxiliary constraint loss, and is dynamically balanced through a course scheduling function. It should be noted that the imitation fidelity, kinematic consistency, and safety constraint losses described later are all specific components of the auxiliary constraint loss, rather than independent overall objectives parallel to the overall training objective.

[0044] in, Indicates the total training loss. This represents the diffusion denoising reconstruction loss. Represents the k-th type of auxiliary loss term; Indicates the first One training round, Indicates the index of the auxiliary loss term. Indicates the first Class Auxiliary Loss in the 1st Dynamic weights across training epochs. The auxiliary loss term is used to implement constraints such as endpoint fidelity, kinematic smoothness, and collision safety, enabling the model to prioritize learning the reachable basic trajectory distribution in the early stages of training, and further optimize smoothness and safety in the later stages of training.

[0045] 1) Imitation fidelity ( To meet the requirements of terminal pose fidelity ( And capturing expert intent, embodiments of the present invention minimize the composite geometric distance consisting of Hausdorff distance, dynamic time warping (DTW), and endpoint MSE: in, This represents the future trajectory predicted by the model. This represents the actual trajectory of the expert for the corresponding sample. This represents the Hausdorff distance between the predicted trajectory and the expert's actual trajectory. This represents the dynamic time-warped distance between the predicted trajectory and the expert's actual trajectory. This indicates the endpoint of the predicted trajectory. This indicates the endpoint of the expert's actual trajectory.

[0046] 2) Kinematic consistency ( To ensure global topology consistency ( In this embodiment of the invention, high-frequency jitter is explicitly penalized. The smoothing loss is achieved by maximizing the adjacent velocity vectors. Cosine similarity to encourage directional continuity: in, Indicates the first The velocity vector between adjacent future waypoints in a segment. Indicates the index of all valid waypoints Find the average. This represents a very small positive number introduced to prevent the denominator from being zero.

[0047] 3) Safety constraints ( (Collision loss term): for any obstacle breakthrough point Safety radius trajectory points Apply soft punishment: 4) Progressive learning strategy: Embodiments of this invention employ... This involves a two-stage learning process. The initial imitation stage (…) Loss weighting is given priority. To establish an effective trajectory manifold, in which, This indicates the preset training round threshold for transitioning from the initial imitation stage to the subsequent refinement stage in the course learning; in the subsequent refinement stage ( The embodiments of the present invention significantly improve and The weights correspond to the weights of the loss term, enhancing the model's focus on trajectory smoothness and collision constraints. This strategy guides the model's transition from basic reachability learning to high-fidelity, collision-free planning refinement oriented towards downstream control.

[0048] The pseudocode for training the conditional diffusion trajectory generation model based on architecture inductive bias is as follows: 1. Input training samples, which include: LiDAR point cloud information at the current moment, robot historical velocity sequence, target coordinates, and expert real trajectory of the corresponding sample.

[0049] 2. Initialize the parameters of the conditional diffusion trajectory generation model based on architecture inductive bias; the model includes a relational inductive bias network, a temporal inductive bias network, and a denoising network.

[0050] 3. For each training round e, iterate through the current batch of training samples.

[0051] 4. For the current sample, randomly sample a diffusion time condition t from the diffusion time step.

[0052] 5. Based on the forward diffusion process, noise is added to the expert's true trajectory to obtain the noisy trajectory at the current diffusion time step t.

[0053] 6. The local LiDAR point cloud P observed at the current moment... t Voxel downsampling is performed, and a local environment graph G=(V,E) is constructed based on the nearest neighbor rule.

[0054] 7. Input the local environment map G into the encoding layer, convolutional layer and pooling layer of the structured scene perception module SSP to obtain node-level scene representation and graph-level LiDAR aggregated features.

[0055] 8. Encode the robot's historical velocity sequence to obtain velocity features.

[0056] 9. By fusing a multilayer perceptron, velocity features are fused with graph-level LiDAR aggregated features to obtain an environmental context vector.

[0057] 10. Encode the target coordinates to obtain target features.

[0058] 11. The initial hidden state is constructed by combining the environmental context vector and the target features.

[0059] 12. In the prediction time domain, make a one-to-one correspondence between prediction steps and future waypoints; for the m-th prediction step, form the context information of the current prediction step based on the previous hidden state and the node-level scene representation.

[0060] 13. Input the current step input, the previous hidden state, the current prediction step context information, the diffusion time condition t, and the target features into the m-th layer graph conditional time-aware GRU to complete the current hidden state update and obtain the temporal hidden state of the corresponding future waypoint.

[0061] 14. After all prediction steps are completed, the hidden state sequence in the future prediction time domain is obtained.

[0062] 15. Input the noisy trajectory into the encoder of the denoising network for feature encoding.

[0063] 16. Input the encoded result and the hidden state sequence into the decoder to obtain the trajectory hiding representation.

[0064] 17. The trajectory hidden representation is sequentially normalized, mapped in a serial fully connected manner, and predicted by the Output MLP to obtain the trajectory increment sequence of all prediction steps.

[0065] 18. Convert the trajectory increment sequence into a future trajectory coordinate sequence by using the cumulative summation operation cumsum.

[0066] 19. When necessary, the final smooth trajectory is obtained by further smoothing through a Bézier curve smoothing layer; during training, the model outputs the predicted future trajectory of the current sample.

[0067] 20. Calculate the diffusion denoising reconstruction loss.

[0068] 21. Calculate the imitation fidelity loss; the imitation fidelity loss is composed of the Hausdorff distance, the dynamic time warping distance, and the endpoint MSE.

[0069] 22. Calculate the kinematic consistency loss; explicitly penalize high-frequency jitter by constraining the directional continuity of adjacent velocity vectors.

[0070] 23. Calculate the safety constraint loss; apply a soft penalty to any trajectory point that exceeds the safety radius of the obstacle point.

[0071] 24. Based on the dynamic weights under the current training round e, the diffusion denoising reconstruction loss and various auxiliary loss terms are weighted and summed to obtain the total training loss.

[0072] 25. Perform backpropagation based on the total training loss and update the model parameters.

[0073] 26. Under the two-stage learning framework, if the current stage is the initial imitation stage, the weight of the corresponding loss term should be increased first in order to establish an effective trajectory manifold.

[0074] 27. If the model is currently in the subsequent refinement stage, increase the corresponding weights of kinematic consistency loss and safety constraint loss to enhance the model's attention to trajectory smoothness and collision constraints.

[0075] 28. Repeat the above process until all training rounds are completed to obtain a trained conditional diffusion trajectory generation model based on architecture inductive bias.

[0076] The effectiveness of the embodiments of the present invention will be verified through experiments below: The experimental data comes from a real outdoor navigation dataset, not a simulated scenario. This dataset was collected by the Husky mobile robot equipped with a Velodyne VLP-16 LiDAR, and the samples include local point cloud observations, velocity history, and target location information.

[0077] Data input format: the last 3 frames of LiDAR and the last 10 frames of velocity history.

[0078] Data output format: a two-dimensional reference trajectory containing 16 waypoints.

[0079] Evaluation method: Offline evaluation in world coordinate system, with the main indicator being Success@2.5m. To further highlight the superiority of the embodiments of the present invention over the prior art, a comparative experiment was conducted for verification. The main comparative models are briefly introduced below: BC: Behavioral Cloning Baseline, a deterministic imitation learning model that directly regresses from observations to output a reference trajectory. It is used to represent the most direct observation-to-trajectory regression path.

[0080] CVAE: Conditional Variational Autoencoder Baseline, a type of raw generative trajectory model. It is used to represent a "generative baseline without post-processing," focusing on comparing the smoothness and geometric quality of the generated trajectory.

[0081] CVAE+bspline: A reference trajectory is first generated using CVAE, and then the generated trajectory is post-processed using B-spline smoothing. This method is not an end-to-end trajectory generation model, but rather serves as a post-processing reference pipeline to compare the degree of improvement in trajectory smoothness and local curvature metrics by relying solely on post-processing methods.

[0082] S2TNet: Sequence modeling baseline, used to provide another type of strong baseline comparison besides the CVAE route, to show that the effect of DRIFT (DRIFT refers to the conditional diffusion trajectory generation model based on architecture inductive bias proposed in the embodiments of this invention) is not only effective for a single generation model.

[0083] Table 1 shows the results of the comparative verification: In Table 1, the meanings of each indicator are as follows: World FDE: Endpoint deviation in the world coordinate system. The smaller the value, the closer the final trajectory endpoint is to the target.

[0084] Success@1m / 2m / 2.5m / 3m: The success rate of the endpoint falling within the corresponding radius threshold. The higher the value, the stronger the target arrival capability.

[0085] Jerk: The intensity of the jerk change in the trajectory; the smaller the value, the smoother the trajectory and the less control impact.

[0086] Inference (ms): The time taken to generate a single trajectory; the smaller the value, the faster the inference speed.

[0087] CurvMean: Average curvature of the trajectory; the smaller the value, the lower the overall steering load.

[0088] CurvMax: Maximum curvature of the trajectory; a smaller value indicates fewer sharp local turns.

[0089] Table 1 Comparison of Verification Results for Key Indicators As shown in Table 1, the conditional diffusion trajectory generation model based on architectural inductive bias proposed in this embodiment of the invention maintains a similar level to generative baselines such as CVAE in terms of World FDE and Success@2.5m metrics. Specifically, DRIFT's World FDE is 2.6904, and its Success@2.5m is 73.57%. Furthermore, compared to CVAE without post-processing, DRIFT reduces Jerk from 525.0221 to 27.1190 and CurvMax from 0.3222 to 0.1010, indicating that it can reduce local oscillations and sharp turns while maintaining target reachability. Compared to the CVAE+bspline post-processing reference pipeline, DRIFT achieves near-perfect mean curvature performance without relying on external smoothing post-processing as the primary mechanism, while maintaining a more direct conditional generation structure. Therefore, Table 1 shows that the technical effects of this invention are mainly reflected in the comprehensive trade-off improvement between target reachability, trajectory smoothness, and local geometric burden.

[0090] The comparison chart of predicted trajectories and actual trajectories from different models is shown below. Figure 2 As shown.

[0091] pass Figure 2 As can be seen, compared with existing models, the DRIFT proposed in this embodiment of the invention reduces local oscillations and sharp geometric corrections while maintaining the overall path structure.

[0092] To further illustrate the impact of the prediction time domain length on the embodiments of the present invention, the trajectory generation effects under different prediction time domains M are compared, and the results are shown in Table 2. Here, Horizon M represents the number of future waypoints included in the generated trajectory; TailErr represents the average error of the last 25% of waypoints in the generated trajectory, used to reflect the stability of the trajectory tail segment.

[0093] Table 2. Sensitivity to Prediction Time Domain Length As shown in Table 2, when the prediction time domain M is 16, the embodiments of the present invention achieve lower World FDE and TailErr while maintaining a higher Success@2.5m. When the prediction time domain is extended to 32 or 64, both the error and inference time increase. These results indicate that the embodiments of the present invention are particularly suitable for generating reference trajectories in the short to medium time domains, and under this setting, they can better balance target reachability, trajectory tail stability, and inference efficiency.

[0094] To further verify the stability of the voxel discretization parameters in the structured scene perception module, the trajectory generation effects under different voxel sizes were compared, and the results are shown in Table 3. Here, Voxel represents the voxel downsampling size of the LiDAR point cloud.

[0095] Table 3. Influence of front-end voxel discretization parameters on trajectory generation effect As shown in Table 3, when the voxel size varies from 0.08 to 0.20, the changes in World FDE, Success@2.5m, Jerk, and CurvMean are relatively small, indicating that the structured scene perception module in this embodiment does not rely on a single special voxel parameter, and the front-end local environment map construction has good parameter stability.

[0096] To further illustrate the friendliness of the generated reference trajectory to downstream control, a simplified tracking playback analysis was performed under the same control conditions in this embodiment of the invention, and the results are shown in Table 4. It should be noted that this analysis is used to compare the tracking difficulty of different reference trajectories under fixed controller conditions, and is not a complete closed-loop real-machine navigation verification. Wherein, Tracking RMSE represents the root mean square error of the tracking process; Final Error represents the final tracking error; Mean|κ| represents the average value of the absolute value of the control curvature; and Sat. Rate represents the proportion of the controller entering saturation.

[0097] Table 4 Comparison of Simplified Tracking Playback Results As shown in Table 4, DRIFT achieved low Tracking RMSE, Final Error, Mean|κ|, and control saturation rate on both the full test set and the accuracy-focused subset. This result indicates that the reference trajectory generated by the embodiments of this invention not only exhibits good target reach and smoothness in open-loop evaluation, but is also easier to stably track under fixed controller playback conditions, thus reducing control curvature burden and control saturation risk.

[0098] This invention also provides a robot trajectory generation system based on diffusion and rule inference. The system generates robot trajectories using a pre-trained conditional diffusion trajectory generation model based on architecture inductive bias. This model includes a relational inductive bias network, a temporal inductive bias network, and a denoising network. The robot trajectory generation system further includes: The data acquisition module is used to acquire the current LiDAR point cloud information, the robot's historical velocity sequence, and the target coordinates; The trajectory generation module is used to perform the following operations: The LiDAR point cloud information, robot historical velocity sequence, and target coordinates are encoded using a relational inductive bias network to obtain LiDAR point cloud information encoding, velocity features, and target features. Graph convolution propagation is performed on the encoded LiDAR point cloud information to obtain node-level scene representations. ;right Graph-level pooling is performed to obtain graph-level LiDAR aggregated features. ; for velocity characteristics Features aggregated with graph-level LiDAR The environment context vector is obtained by fusion. ; Under the joint constraints of global environment context vector, target features, and node-level scene representation, in the prediction time domain, the graph conditional time-aware GRU of the temporal inductive bias network is sequentially expanded into layers 1 to m, and the temporal hidden state of the future trajectory is recursively generated along the prediction time domain. , … ; Temporally hidden states are processed using a denoising network. , … The noisy input trajectory is processed to obtain the robot trajectory.

[0099] It is understood that the robot trajectory generation method based on diffusion and rule inference provided in this embodiment of the invention corresponds to the robot trajectory generation system based on diffusion and rule inference described above. The explanations, examples, and beneficial effects of the relevant content can be referred to the corresponding content in the robot trajectory generation system based on diffusion and rule inference, and will not be repeated here.

[0100] This invention also provides a computer-readable storage medium storing a computer program for generating robot trajectories based on diffusion and rule inference, wherein the computer program causes a computer to execute the robot trajectory generation method based on diffusion and rule inference as described above.

[0101] This invention also provides an electronic device, including: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the robot trajectory generation method based on diffusion and rule inference as described above.

[0102] In summary, compared with existing technologies, it has the following beneficial effects: 1. This invention proposes a novel conditional diffusion trajectory generation model based on architectural inductive bias. This framework does not rely solely on a loss function for result constraints, but decomposes the complex trajectory generation task into two complementary technical processes: First, based on local environment graph construction, graph convolutional propagation, and graph-level feature convergence, it achieves reasoning about the overall traversable structure and obstacle topology within the current local environment graph. Second, based on node-level scene representation, continuous injection of target features, and step-by-step context updates, it achieves fine-grained local maneuver control of obstacle avoidance, turning, and terminal convergence behaviors at each future waypoint. This solves the technical problem that existing diffusion-based trajectory generation methods cannot dynamically adjust environmental information, ensuring high-fidelity trajectory generation in complex unstructured environments. Simultaneously, this invention designs a graph conditional time-aware GRU (GTGRU), which, by introducing a temporal attention bias, solves the terminal drift problem caused by the reliance on static conditional encoding in standard diffusion planners.

[0103] 2. This invention presents an SSP module based on a graph neural network (GNN), which uses relational inductive bias to encode disordered, discrete LiDAR point clouds into a latent space scene graph with topological consistency. This technology can explicitly capture the spatial availability of the environment, provide globally consistent rule guidance for the diffusion process, and pre-suppress high jerk factors that cause motion jitter at the architectural level. This reduces high-frequency jitter and abrupt changes in the trajectory, lowers the trajectory abruptness level, improves the tracking stability and execution smoothness of the downstream controller on the reference trajectory, reduces the risk of overshoot during the robot's terminal approach phase, and enhances its safety in fine maneuvers in confined spaces or complex obstacle environments.

[0104] 3. This embodiment of the invention employs a unique decoupling strategy in the feature fusion stage: global environment context vector ( The system deliberately excludes target point information, focusing instead on the accessibility of the environment (satisfying global topology constraints). Meanwhile, target guidance information is continuously and dynamically injected through the GTGRU's gating mechanism. This design ensures that the perception and generation modules each perform their respective functions, effectively mitigating potential conflicts between obstacle avoidance and target approach.

[0105] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0106] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A robot trajectory generation method based on diffusion and rule inference, characterized in that, Robot trajectories are generated by a pre-trained conditional diffusion trajectory generation model based on architecture inductive bias, wherein the conditional diffusion trajectory generation model includes a relational inductive bias network, a temporal inductive bias network, and a denoising network. The robot trajectory generation method includes: Obtain the current LiDAR point cloud information, robot historical velocity sequence, and target coordinates; The LiDAR point cloud information, robot historical velocity sequence, and target coordinates are encoded using a relational inductive bias network to obtain LiDAR point cloud information encoding, velocity features, and target features. Graph convolution propagation is then applied to the LiDAR point cloud information encoding to obtain node-level scene representations. Graph pooling is then applied to the node-level scene representations to obtain graph-level LiDAR aggregated features. Finally, the velocity features and graph-level LiDAR aggregated features are fused to obtain the environmental context vector. Under the joint constraints of global environment context vector, target features, and node-level scene representation, in the prediction time domain, the graph conditional time-aware GRU of the temporal inductive bias network is sequentially expanded into layers 1 to m, and the temporal hidden state of the future trajectory is recursively generated along the prediction time domain. , … ,include: Node-level scene representation Together with the current recursive state, they form the context information of the current prediction step. For the first prediction step, based on the initial hidden state... Node-level scene representation Forming the first prediction step context information For the m-th prediction step, based on the hidden state of the previous prediction step... Node-level scene representation Forming context information for the current prediction step ; where the initial hidden state The expression is: in, Represents the environment context vector The mapping weight matrix to the initial hidden state space. Representing target features The mapping weight matrix to the initial hidden state space. Indicates the initial hidden state; For the first future waypoint, input the current step. Initial hidden state First prediction step context information diffusion time condition t and target characteristics The first-layer graph conditional time-aware GRU is input together, and the first temporal hidden state is obtained through updating. ; For the m-th future path point, input the current step. Previous hidden state Current prediction step context information diffusion time condition t and target characteristics Input the conditional time-aware GRU of the m-th layer graph together to complete the current hidden state update; Temporally hidden states are processed using a denoising network. , … The noisy input trajectory is processed to obtain the robot trajectory.

2. The robot trajectory generation method based on diffusion and rule inference as described in claim 1, characterized in that, The relation induction bias network includes a structured scene perception module (SSP) based on a GNN and a fusion multilayer perceptron. The structured scene perception module includes an encoding layer, a convolutional layer, and a pooling layer; The encoding layer performs raw conditional encoding on the current LiDAR point cloud information, the robot's historical velocity sequence, and the target coordinates, respectively, to obtain the LiDAR point cloud information encoding and velocity features. and target features ; Convolutional layers perform graph convolution propagation on the LiDAR point cloud information encoding to obtain node-level scene representations. ; Pooling layer Graph-level pooling is performed to obtain graph-level LiDAR aggregated features. ; Fusion of speed features from multilayer perceptrons Features aggregated with graph-level LiDAR The elements are fused to form an environmental context vector. .

3. The robot trajectory generation method based on diffusion and rule inference as described in any one of claims 1 to 2, characterized in that, The denoising network includes an encoder, a decoder, and an output layer multilayer perceptron; the output layer multilayer perceptron includes a layer normalization layer, a serial fully connected layer, a cumulative summation layer, and a Bézier curve smoothing layer. Among them, the encoder outputs the noisy trajectory of the input. Feature encoding is performed, and the decoder receives and processes the encoding result and the hidden state sequence. , … The trajectory hiding representation is obtained. ; Layer normalization layer for trajectory hiding representation After processing, a stable hidden representation is obtained, denoted as . ; Serial fully connected layers for stable hidden representation Perform sequence mapping to obtain the trajectory increment sequence of all prediction steps; After obtaining the trajectory increments of all prediction steps, the future trajectory coordinate sequence is calculated through a cumulative summation layer; The Bézier curve smoothing layer smooths the future trajectory coordinate sequence to obtain the robot trajectory.

4. The robot trajectory generation method based on diffusion and rule inference as described in any one of claims 1 to 2, characterized in that, The loss function of the pre-trained conditional diffusion trajectory generation model based on architecture inductive bias during training includes: in, Indicates the total training loss; This represents the loss during diffusion denoising and reconstruction. Represents the k-th type of auxiliary loss term; Indicates the first One training round, Indicates the index of the auxiliary loss term. Indicates the first Class Auxiliary Loss in the 1st Dynamic weights across training rounds.

5. The robot trajectory generation method based on diffusion and rule inference as described in claim 4, characterized in that, The auxiliary loss terms include imitation fidelity loss, kinematic consistency loss, and safety constraint loss; Among them, the loss of imitation fidelity include: in, This represents the future trajectory predicted by the model. This represents the actual trajectory of the expert for the corresponding sample. This represents the Hausdorff distance between the predicted trajectory and the expert's actual trajectory. This represents the dynamic time-warped distance between the predicted trajectory and the expert's actual trajectory. This indicates the endpoint of the predicted trajectory. Indicates the endpoint of the expert's actual trajectory; Loss of kinematic consistency include: in, Indicates the first The velocity vector between adjacent future waypoints in a segment. , Indicates the index of all valid waypoints Find the average. This represents a very small positive number introduced to prevent the denominator from being zero; Safety constraint loss include: For any breakthrough point Safety radius trajectory points Apply soft punishment: 。 6. The robot trajectory generation method based on diffusion and rule inference as described in claim 5, characterized in that, The pre-trained conditional diffusion trajectory generation model based on architectural inductive bias is trained through... The course will be implemented in two phases, specifically including: when This is the initial imitation stage, aiming to improve... The weights corresponding to the loss terms; when In order to improve the efficiency in the subsequent refining stage. and The weights corresponding to the loss terms; where, This indicates the preset training round threshold for switching from the initial imitation stage to the subsequent refinement stage in the course learning process.

7. A robot trajectory generation system based on diffusion and rule inference, characterized in that, The robot trajectory generation system generates robot trajectories through a pre-trained conditional diffusion trajectory generation model based on architecture inductive bias. The conditional diffusion trajectory generation model includes a relational inductive bias network, a temporal inductive bias network, and a denoising network. The robot trajectory generation system includes: The data acquisition module is used to acquire the current LiDAR point cloud information, the robot's historical velocity sequence, and the target coordinates; The trajectory generation module is used to perform the following operations: The LiDAR point cloud information, robot historical velocity sequence, and target coordinates are encoded using a relational inductive bias network to obtain LiDAR point cloud information encoding, velocity features, and target features. Graph convolution propagation is then applied to the LiDAR point cloud information encoding to obtain node-level scene representations. Graph pooling is then applied to the node-level scene representations to obtain graph-level LiDAR aggregated features. Finally, the velocity features and graph-level LiDAR aggregated features are fused to obtain the environmental context vector. Under the joint constraints of global environment context vector, target features, and node-level scene representation, in the prediction time domain, the graph conditional time-aware GRU of the temporal inductive bias network is sequentially expanded into layers 1 to m, and the temporal hidden state of the future trajectory is recursively generated along the prediction time domain. , … ,include: Node-level scene representation Together with the current recursive state, they form the context information of the current prediction step. For the first prediction step, based on the initial hidden state... Node-level scene representation Forming the first prediction step context information For the m-th prediction step, based on the hidden state of the previous prediction step... Node-level scene representation Forming context information for the current prediction step ; where the initial hidden state The expression is: in, Represents the environment context vector The mapping weight matrix to the initial hidden state space. Representing target features The mapping weight matrix to the initial hidden state space. Indicates the initial hidden state; For the first future waypoint, input the current step. Initial hidden state First prediction step context information diffusion time condition t and target characteristics The first-layer graph conditional time-aware GRU is input together, and the first temporal hidden state is obtained through updating. ; For the m-th future path point, input the current step. Previous hidden state Current prediction step context information diffusion time condition t and target characteristics Input the conditional time-aware GRU of the m-th layer graph together to complete the current hidden state update; Temporally hidden states are processed using a denoising network. , … The noisy input trajectory is processed to obtain the robot trajectory.

8. A computer-readable storage medium, characterized in that, It stores a computer program for generating robot trajectories based on diffusion and rule inference, wherein the computer program causes a computer to execute the robot trajectory generation method based on diffusion and rule inference as described in any one of claims 1 to 6.

9. An electronic device, characterized in that, include: One or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs including methods for performing the robot trajectory generation method based on diffusion and rule inference as described in any one of claims 1 to 6.

Images