Unmanned aerial vehicle adaptive motion planning method and device for different environments

Abstract

The application belongs to the technical field of unmanned aerial vehicle planning, and relates to an unmanned aerial vehicle adaptive motion planning method and device for different environments. The method comprises the following steps: constructing an initial overall planning network according to a deep observation feature and a state observation feature; constructing an initial objective function according to the initial overall planning network; obtaining scene domain pictures of different environments, sampling, obtaining an initial objective function of each sample, then obtaining an expected domain risk of each scene domain, introducing a domain weight vector, calculating a weighted domain risk of all scene domains, and obtaining a worst domain objective function; taking the worst domain objective function as a constraint to train the initial overall planning network, and obtaining an optimized overall planning network; obtaining scene domain pictures of an unknown environment, inputting the optimized overall planning network, outputting an end state parameter constraint, and constructing a motion trajectory of the unmanned aerial vehicle. The application can realize unmanned aerial vehicle adaptive motion planning.

Description

Technical Field

[0001] This application relates to the field of unmanned aerial vehicle (UAV) planning technology, and in particular to UAV adaptive motion planning methods and apparatus for different environments. Background Technology

[0002] As a tool demonstrating great potential in complex scenarios such as disaster relief, environmental monitoring, and urban logistics, unmanned aerial vehicles (UAVs) excel in safe, efficient, and autonomous motion planning. However, developing motion planning algorithms with strong generalization capabilities and stable operation in various environments remains a significant challenge.

[0003] In recent years, the high dynamism and complex spatial structure of the real world environment have prompted people to increasingly focus on improving the real-time performance and intelligence of drone motion planning, and thus several representative planning paradigms have been developed.

[0004] The first category of methods includes classic deterministic methods, such as: The algorithms include Dijkstra's algorithm and the Artificial Potential Field (APF) algorithm. These methods are known for their stability in static scenes, as they are based on explicit environment models; however, they face significant challenges in partially observable or unknown environments, primarily due to their high sensitivity to perceptual errors, occlusion, and dynamic changes.

[0005] The second category includes sampling and stochastic planning methods, such as Probabilistic Roadmap (PRM), Rapid Exploratory Random Tree (RRT), and their rapidly expanding variants. Compared to classical deterministic methods, these methods have powerful mapping and search capabilities in high-dimensional spaces; however, their performance is heavily dependent on the environment structure and hyperparameters, which often leads to performance degradation in sparse feasible regions or narrow channels.

[0006] The third category includes heuristic optimization methods and swarm intelligence algorithms. Similarly, these methods can also explore complex search spaces; however, their high computational cost makes them difficult to meet the real-time requirements of online drone motion planning scenarios.

[0007] Besides the methods mentioned above, machine learning (ML) methods have stood out due to their superior expressive power in handling high-dimensional perceptual inputs, modeling environmental structures, and searching feasible trajectories, attracting widespread research attention. Among them, supervised learning and reinforcement learning, by directly predicting feasible trajectories or policies from visual or state inputs, have become important means of achieving the fusion of perception and planning. However, despite the significant advantages of machine learning methods, they still face the challenge of generalization traps in practical applications: these methods typically rely on large-scale training environments with relatively fixed structures, and the learned policies are often "environment-specific" rather than "environment-invariant structures" that can be shared across scenarios. When the test environment deviates from the training distribution in terms of topology, obstacle layout, or dynamic characteristics, the model performance often drops significantly or even fails completely. Summary of the Invention

[0008] Therefore, it is necessary to provide a method and apparatus for adaptive motion planning of UAVs in different environments to address the above-mentioned technical problems, which can realize adaptive motion planning of UAVs.

[0009] Adaptive motion planning methods for UAVs in different environments include:

[0010] Based on the characteristics of depth observation and state observation, an initial overall planning network is constructed; Based on the initial overall planning network, construct the initial objective function; Scene domain images from different environments are acquired and sampled to obtain the initial objective function for each sample. Based on the initial objective function for each sample, the expected domain risk for each scene domain is obtained. A domain weight vector is introduced for the expected domain risk of each scene domain, and the weighted domain risk of all scene domains is calculated to obtain the worst-case domain objective function. Using the worst-case domain objective function as a constraint, the domain weight vector is updated to train the initial global planning network until the preset conditions are met, thus obtaining the optimized global planning network. The system acquires scene domain images of unknown environments, inputs them into an optimized overall planning network, and outputs end-point state parameter constraints to construct the motion trajectory of the UAV and achieve adaptive motion planning.

[0011] In one embodiment, an initial global planning network is constructed based on depth observation features and state observation features, including: Based on the depth observation features and state observation features, feature streams are extracted separately and connected along the channel dimension to obtain the fused input; The fused input is projected onto the latent representation, and multiple expert networks are introduced to obtain the output of each expert network. The latent representation is introduced into a router to obtain expert logical values, and a sparse gating method is used to establish an expert set. Normalization is performed within the expert set to obtain the gating weights. Based on the output of each expert network and the gating weights, the expert weighted sum is obtained; The expert weights are then mapped back to the original channel dimension to obtain the mapped input; Based on the mapping input, construct the initial global planning network.

[0012] In one embodiment, the fused input is projected onto a latent representation, and multiple expert networks are introduced to obtain the output of each expert network, including: ; ; In the formula, For potential representation, For feature fusion, To integrate input, For the first The output of an expert network, For the first A network of experts, The number of expert networks.

[0013] In one embodiment, constructing an initial global planning network based on the mapping input includes: ;

[0014] In the formula, For the first The end state parameter constraints of the candidates, for The corresponding predicted cost, The number of candidate trajectories. To share the prediction head, For mapping input, For the initial overall planning network, for The depth observation characteristics at any given time for State observation characteristics at any given time For projection head, For hybrid expert modules, For feature fusion, For deep feature extractors, To spatialize state observations through coordinate transformation and grid alignment.

[0015] In one embodiment, an initial objective function is constructed based on the initial overall planning network, including: Based on the initial global planning network, the end state parameter constraints are obtained, and the end state parameter constraints are parameterized in the world coordinate system to obtain the parameterized end state parameter constraints. The starting state is obtained, and combined with the parameterized ending state parameter constraints, an analytical polynomial generator is used to construct candidate trajectories. Construct differentiable cost labels based on candidate trajectories; Based on the differentiable cost label, trajectory loss and evaluation loss are established, and an initial objective function is constructed.

[0016] In one embodiment, based on the differentiable cost label, trajectory loss and evaluation loss are established, and an initial objective function is constructed, including: ; in, ; ; In the formula, Let the initial objective function be... For trajectory loss weighting coefficients, For trajectory loss, To evaluate the loss weighting coefficient, To assess the loss; The number of candidate trajectories. Differentiable cost label; for, for The corresponding predicted cost, For the first The end state parameter constraints of the candidates, To stop gradient propagation operator.

[0017] In one embodiment, the expected domain risk for each scenario domain is obtained based on the initial objective function for each sample, including: ; In the formula, For the first Expected risks in each scenario domain. For the data distribution from the k-th scene Data distribution for samples Expectations The initial objective function for each sample, The number of scene domains.

[0018] In one embodiment, a domain weight vector is introduced for the expected domain risk of each scenario domain, and the weighted domain risk of all scenario domains is calculated to obtain the worst-case domain objective function, including: ; In the formula, For the model's learnable parameters, For the domain weight vector, Let K be the weight set of the scene domains. Weighted domain risk for all scenario domains For the first Domain weight vectors for each scene domain For the first Expected domain risks in each scenario domain.

[0019] In one embodiment, the domain weight vector is updated using the worst-case domain objective function as a constraint to train the initial global planning network until a preset condition is met, resulting in an optimized global planning network, including: The worst-case domain objective function is used as the constraint for the initial global programming network to estimate the empirical risk, and the domain weight vector is updated using the exponential gradient ascent method. Based on the updated domain weight vector, the batch-level DRO loss is updated to train the initial global planning network until the preset conditions are met, resulting in an optimized global planning network.

[0020] An adaptive motion planning device for UAVs in different environments, employing the aforementioned adaptive motion planning method for UAVs in different environments, includes: The first module is used to construct an initial overall planning network based on the depth observation characteristics and state observation characteristics; The second module is used to construct the initial objective function based on the initial overall planning network; The third module is used to acquire scene domain images of different environments, sample them, and obtain the initial objective function for each sample; based on the initial objective function for each sample, the expected domain risk for each scene domain is obtained; a domain weight vector is introduced for the expected domain risk of each scene domain, and the weighted domain risk of all scene domains is calculated to obtain the worst-case domain objective function. The fourth module is used to update the domain weight vector according to the worst-case domain objective function in order to train the initial overall planning network until the preset conditions are met, and then obtain the optimized overall planning network. The fifth module is used to acquire scene domain images of unknown environments, input them into the overall planning network for optimization, and output end-point state parameter constraints to construct the UAV's motion trajectory and achieve adaptive motion planning.

[0021] The aforementioned UAV adaptive motion planning method and device for different environments, designed for the UAV local planning problem based on visual depth observation, employs a robust dual-track mechanism. By parallel processing model architecture design and training algorithm optimization, it can mitigate the bias of test distribution under multi-environment training, enabling the UAV to maintain reliable and safe planning behavior when environmental geometry (different environments have different geometric features, corresponding to different state observation features) and observation features (corresponding to different depth observation features) change. This achieves cross-domain generalization and stable execution of short-term planning based on UAV vision, improving stable cross-scene learning ability, cross-environment generalization ability, and cross-domain robustness. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating an adaptive motion planning method for UAVs in different environments, as shown in one embodiment. Figure 2 This is a schematic diagram of the architecture of an adaptive motion planning method for UAVs in different environments in one embodiment; Figure 3 This is a trajectory diagram of a drone adaptive motion planning method for different environments in a wall scene, as shown in a specific embodiment. Figure 4 Selecting an interaction area in a specific embodiment A diagram illustrating the key moments; Figure 5 Selecting an interaction area in a specific embodiment A diagram illustrating the key moments; Figure 6 Selecting an interaction area in a specific embodiment A diagram illustrating the key moments; Figure 7 This is the original input depth image at key moment a in a specific embodiment; Figure 8 A pseudo-color depth visualization image of key moment a in a specific embodiment; Figure 9 Here is a Grad-CAM attention image for key moment a in a specific embodiment; Figure 10 This is the original input depth image at key moment c in a specific embodiment; Figure 11 A pseudo-color depth visualization image of key moment c in a specific embodiment; Figure 12 This is a Grad-CAM attention image at key moment c in a specific embodiment; Figure 13 This is the original input depth image at key moment e in a specific embodiment; Figure 14 A pseudo-color depth visualization image of key moment e in a specific embodiment; Figure 15 Here is a Grad-CAM attention image for key moment e in a specific embodiment; Figure 16 This is the original input depth image at a key moment g in a specific embodiment; Figure 17 A pseudo-color depth visualization image of key moment g in a specific embodiment; Figure 18 This is a Grad-CAM attention image of key moment g in a specific embodiment; Figure 19 This is the original input depth image at a key moment h in a specific embodiment; Figure 20 A pseudo-color depth visualization image of key moment h in a specific embodiment; Figure 21 This is a Grad-CAM attention image at a key moment h in a specific embodiment; Figure 22 This is an overall environmental layout diagram of a forest environment in a specific embodiment; Figure 23 This is a top view of the planning and execution trajectory from the starting point to the end point in a forest environment, as shown in a specific embodiment. Figure 24 This is an overall environmental layout diagram of the columnar environment in a specific embodiment; Figure 25 This is a top view of the planning and execution trajectory from the starting point to the end point in a columnar environment in a specific embodiment; Figure 26 This is a structural block diagram of an adaptive motion planning device for unmanned aerial vehicles (UAVs) in different environments, as shown in one embodiment. Figure 27 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without inventive effort are within the scope of protection of this application.

[0024] Furthermore, the use of terms such as "first" and "second" in this application is for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of those features. In the description of this application, "multiple sets" means at least two sets, such as two sets, three sets, etc., unless otherwise explicitly specified.

[0025] In this application, unless otherwise expressly specified and limited, the terms "connection," "fixed," etc., should be interpreted broadly. For example, "fixed" can mean a fixed connection, a detachable connection, or an integral part; it can mean a mechanical connection, an electrical connection, a physical connection, or a wireless communication connection; it can mean a direct connection or an indirect connection through an intermediate medium; it can mean the internal communication of two elements or the interaction between two elements, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0026] Furthermore, the technical solutions of the various embodiments of this application can be combined with each other, but only if they are based on the ability of those skilled in the art to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by this application.

[0027] This application provides an adaptive motion planning method for UAVs in different environments, such as... Figure 1 The flowchart shown, in one embodiment, includes: Step 101: Construct an initial overall planning network based on the depth observation characteristics and state observation characteristics.

[0028] Specifically: Based on the depth observation features and state observation features, feature streams are extracted separately and connected along the channel dimension to obtain the fused input; The fused input is projected onto the latent representation, and multiple expert networks are introduced to obtain the output of each expert network. The latent representation is introduced into a router to obtain expert logical values, and a sparse gating method is used to establish an expert set. Normalization is performed within the expert set to obtain the gating weights. Based on the output of each expert network and the gating weights, the expert weighted sum is obtained; The expert weights are then mapped back to the original channel dimension to obtain the mapped input; Based on the mapping input, construct the initial global planning network.

[0029] More specifically: Based on the depth observation features and the state observation features in the body coordinate system, feature flows are extracted separately and connected along the channel dimension to obtain the fused input: ; ; In the formula, For deep observation features, For deep feature extractors, for The depth observation characteristics at any given time The feature flow is a feature of state observation characteristics. To spatialize state observations through coordinate transformation and grid alignment, for State observation characteristics at any given time For input fusion; The fused input is projected onto a latent representation, and multiple expert networks are introduced, each learning a different nonlinear transformation in a shared latent space to capture domain-dependent variations in the fusion pattern caused by structural differences, resulting in the output of each expert network: ; ; in, ; In the formula, For potential representation, For feature fusion, To integrate input, For the first The output of an expert network, For the first A network of experts, The number of expert networks; By introducing the latent representation into a router to achieve adaptive structure selection, expert logical values ​​are obtained: ; In the formula, For expert logic values, For routers, For potential representation, For the real number field; Based on the expert logic values, a TopK sparse gating method is used to establish an expert set, and the gating weights are obtained by normalization within the expert set. ; In the formula, For the gathering of experts, The operation is to select the first few (or more) maximum values ​​from the experts. For expert logic values, To select the number of experts to activate (in this application) Specifically, it can be selected as 2). Based on the output of each expert network and the gating weights, the weighted sum of the activated experts is obtained: ; In the formula, Weighted sum by experts, For the number of expert networks, For gating weights, For the first The output of an expert network; Application of projection head The output projection maps the expert weighted sum back to the original channel dimension (i.e., the baseline feature dimension) to preserve the interface. Without changing the input, we get the mapped input: ; In the formula, For mapping input, For projection head, Weighted sum by experts; Based on the mapping input, construct the initial global planning network:

[0030] In the formula, For the first The end state parameter constraints of the candidates, for The corresponding predicted cost, The number of candidate trajectories. To share the prediction head, For mapping input, For the initial overall planning network, for The depth observation characteristics at any given time for State observation characteristics at any given time For projection head, For hybrid expert modules, For feature fusion, For deep feature extractors, To spatialize state observations through coordinate transformation and grid alignment.

[0031] In this step, a structure-aware hybrid expert (MoE) module was designed to provide conditional fusion mapping for scene domains with different structures. This addresses the problem that "different scene domains have significant differences in the distribution of geometric configuration and traversable free space. Relying on a single fixed fusion mapping will force feature representations to compromise between different scene domains, which will lead to performance degradation when the distribution shifts." This module comprises a shared representation encoder, a multi-expert fusion network, and a shared task decoder. The shared representation encoder extracts depth and state observations and spatially aligns them to concatenate along the channel dimension, forming a unified fusion input. The multi-expert fusion network introduces multiple expert networks whose outputs provide multiple candidate fusion transformations within a shared latent space. It performs sample-by-sample adaptive fusion via sparse routing (i.e., structure-aware routing with Top-k sparse activations, including routers and Top-k sparse gating), outputting an expert-weighted sum of activations to adapt to different environments. The shared task decoder maps the fused representation (processed latent representation) back to the baseline feature dimension and connects it to the same prediction head, thereby improving the robustness of fusion to cross-domain structural variations without altering the output interface. Cross-domain structural variations (i.e., different scene domains) are primarily absorbed by the multi-expert fusion network, while the shared task decoder remains shared and operates within a unified output space. This design enhances the robustness of the fused representation (processed latent representation) to distribution shifts during joint training and testing across multiple environments, thereby improving the reliability of candidate generation and cost-based ranking.

[0032] It should be noted that how to normalize within the expert set to obtain the gating weights is an existing technology and will not be elaborated here.

[0033] Step 102: Construct the initial objective function based on the initial overall planning network.

[0034] Specifically: Based on the initial global planning network, the end state parameter constraints are obtained, and the end state parameter constraints are parameterized in the world coordinate system to obtain the parameterized end state parameter constraints. The starting state is obtained, and combined with the parameterized ending state parameter constraints, an analytical polynomial generator is used to construct candidate trajectories. Construct differentiable cost labels based on candidate trajectories; Based on the differentiable cost label, trajectory loss and evaluation loss are established, and an initial objective function is constructed.

[0035] More specifically: Based on the initial overall network planning, a candidate set parallel evaluation paradigm is adopted, in which the network outputs results simultaneously during a single forward inference. The candidate terminal state parameter constraints and their corresponding prediction costs are determined to satisfy real-time closed-loop requirements. ; In the formula, For the first End-state parameter constraints for each candidate. for The corresponding prediction cost (the smaller the better). The number of candidate trajectories. For the initial overall planning network, for The depth observation characteristics at any given time for State observation characteristics at any given time; The end-state parameter constraints are parameterized in the world coordinate system to obtain the parameterized end-state parameter constraints: ; In the formula, For the first The parameterized end-state parameter constraints of the candidate, To define symbols, For the first The target position parameters in the world coordinate system for each candidate terminal state. For the first The target velocity parameters in the world coordinate system for each candidate terminal state. For the first The target acceleration parameters in the world coordinate system for each candidate terminal state. For the real number field; The starting state is obtained, and combined with parameterized ending state parameter constraints, an analytic polynomial generator is used to construct candidate trajectories to establish an analytic mapping from the ending state to the executable trajectory, ensuring the continuity of position, velocity, and acceleration, and meeting the trackability and executable requirements of the downstream controller. ; in, ; In the formula, Candidate trajectories, To parse polynomial generators, This is the initial state. To parameterize the end-state parameter constraints, Time domain planning for trajectory, The starting position in the world coordinate system. The starting velocity in the world coordinate system. The acceleration at the starting point in the world coordinate system; For each candidate trajectory, construct a differentiable cost label: ; In the formula, For differentiable cost labels, As a term to measure the smoothness of candidate trajectories, As a measure of the safety of candidate trajectories, To measure the target-guided parameters of the candidate trajectory; Based on the differentiable cost label, establish the trajectory loss and evaluation loss: ; ; In the formula, For trajectory loss, To assess the loss; The number of candidate trajectories. Differentiable cost label; To smooth the L1 loss function, for The corresponding predicted cost, For the first The end state parameter constraints of the candidates, To stop gradient propagation operator; Based on the trajectory loss and evaluation loss, construct the initial objective function: ; In the formula, Let the initial objective function be... For trajectory loss weighting coefficients, For trajectory loss, To evaluate the loss weighting coefficient, To evaluate the loss.

[0036] In this step: Establish trajectory loss This can minimize the true cost of generating candidates, thereby optimizing the terminal prediction. Quality; Establish an evaluation system for losses ,by Supervisory rating head ,use Regression, and apply to the label To decouple the gradient propagation of score regression and terminal branches, and to enforce the consistency of cost prediction; the two types of loss provide complementary supervision.

[0037] Step 103: Obtain scene domain images of different environments, sample them, and obtain the initial objective function for each sample; based on the initial objective function for each sample, obtain the expected domain risk for each scene domain; introduce a domain weight vector into the expected domain risk for each scene domain, calculate the weighted domain risk for all scene domains, and obtain the worst-case domain objective function.

[0038] Specifically: Obtain scene domains for different environments The images are sampled, and the initial objective function for each sample is obtained based on the constructed initial objective function: ; In the formula, for The initial objective function for each sample in the training instances sampled from the scene domain. for Trajectory loss for each sample in the dataset. for The evaluation loss for each sample in the dataset; Based on the initial objective function for each sample, the expected domain risk for each scenario domain is obtained: ; In the formula, For the first Expected risks in each scenario domain. For the data distribution from the k-th scene Data distribution for samples Expectations The initial objective function for each sample, The number of scene domains; For the expected domain risk of each scenario domain, a domain weight vector is introduced to avoid the problem of "directly minimizing the average risk causing errors to be masked in the difficult domain": ; in, ; In the formula, For the domain weight vector, Let K be the weight set of the scene domains. To define symbols, Let K be a K-dimensional non-negative real vector space. For the number of scene domains, For the first Domain weight vectors for each scene domain; Based on the domain weight vector, the weighted domain risk of all scenario domains is calculated to obtain the worst-case domain objective function (where the inner maximization concentrates the weights on the high-risk domain, thus ensuring that the outer minimization maintains the same robustness as the worst-case domain): ; in, ; In the formula, For the model's learnable parameters, Weighted domain risk for all scenario domains For the number of scene domains, For the first Domain weight vectors for each scene domain For the first Expected domain risks in each scenario domain.

[0039] In this step: the worst-case domain objective function is a domain-level training objective based on Distributed Robust Optimization (DRO). This objective dynamically focuses on the most difficult environment, keeping training performance consistent with the worst-case domain, improving cross-domain robustness, and addressing the problem that "existing methods often favor optimization in simpler domains with more stable gradients, while insufficient optimization in difficult domains leads to a significant performance drop when the distribution changes during testing (unseen structures appear)." In addition, the worst-case domain objective function, combined with structure-aware MoE, can improve the stability of candidate generation and cost-based ranking under distribution changes during testing, thereby maintaining a larger safety margin.

[0040] Step 104: Using the worst-case domain objective function as a constraint, update the domain weight vector to train the initial global planning network until the preset conditions are met, thus obtaining the optimized global planning network.

[0041] Specifically: The worst-case domain objective function is used as the constraint for the initial global programming network to estimate the empirical risk, and the domain weight vector is updated using the exponential gradient ascent method. Based on the updated domain weight vector, the batch-level DRO loss is updated to train the initial global planning network until the preset conditions are met, resulting in an optimized global planning network.

[0042] More specifically: Using the worst-case domain objective function as the initial constraint for the global programming network, samples are grouped according to their domain identifiers in each mini-batch to estimate empirical risk, and the domain weight vector is updated using the exponential gradient ascent method. ; In the formula, for The first moment Domain weight vectors for each scene domain for The first moment Domain weight vectors for each scene domain It is a natural exponential function. To control the temperature parameter of the weighted concentration, For experience-based risk, For the first One scenario, Let be the weight corresponding to the j-th scene at time t. For the experience risk of the j-th scenario; Update based on the updated domain weight vector The batch-level DRO loss of gradient descent is used to train the initial global planning network until a preset condition is met, at which point convergence is determined, resulting in an optimized global planning network. ; In the formula, For batch-level DRO loss, Risk based on experience.

[0043] In this step: since the trajectory selection is jointly determined by candidate generation and cost ranking, prediction errors can lead to incorrect selections and collision risks. Therefore, the focus is on training the planning strategy. This enables it to accurately generate candidate terminal states and reliably predict the corresponding costs for selecting the optimal trajectory. In the multi-environment joint training setting, the training data comes from... A scene domain with different structures DRO performs adaptive domain-level reweighting to prevent error accumulation in difficult domains and shifts the optimization objective from average-case control to worst-case domain control performance.

[0044] It should be noted that the methods for training the initial global planning network based on the updated batch-level DRO loss, determining whether convergence has been achieved, and obtaining the optimized global planning network are all existing technologies and will not be elaborated here.

[0045] Step 105: Obtain scene domain images of the unknown environment, input them into the optimized overall planning network, and output end state parameter constraints to construct the motion trajectory of the UAV and achieve adaptive motion planning.

[0046] Specifically: Acquire scene domain images of unknown environments, extract depth observation features and state observation features, input them into the overall planning network for optimization, and output end state parameter constraints; The starting state is obtained, and the motion trajectory of the UAV is constructed based on the constraints of the ending state parameters to achieve adaptive motion planning.

[0047] In this step: how to acquire scene domain images of unknown environments, how to extract depth observation features and state observation features, how to obtain the starting state, and how to construct the motion trajectory of the UAV are all existing technologies and will not be elaborated here.

[0048] In this embodiment, an Adaptive Environment-Robust Optimization for Reliable Vision-Based UAVNavigation (AERO) method is designed. This is a unified framework that improves the cross-environment generalization capability of vision-based short-time domain planning from two complementary perspectives: model architecture and training optimization. Regarding model architecture, a "Structure-Aware Hybrid Expert" (MoE) module for deep-state fusion is introduced, including an expert network and sparse routing. The sparse routing adaptively selects expert transformations to capture domain-invariant structures and domain-specific variations while maintaining the predictor head and inference interface unchanged. For training optimization, a worst-domain-distributionally-robust optimization (DRO) objective function is adopted, which dynamically emphasizes the difficult domain during multi-environment training, thus explicitly keeping the learning consistent with the worst-case performance.

[0049] like Figure 2 As shown, a structure-aware hybrid expert (MoE) module is introduced in the multimodal fusion process of deep observation and self-state observation. This module uses sparse routing to adaptively select the expert transform (choosing the best expert network), thereby explicitly capturing structural differences in different scenarios. Simultaneously, the prediction head and output interface (for the network, the input is always an image, and the output is always a trajectory) remain unchanged, improving cross-environment generalization ability without increasing the complexity of the inference interface. During multi-environment joint training, a worst-case domain objective function is adopted. By dynamically increasing the loss weights, hard-domain samples are identified and emphasized online, aligning optimization with the worst-case domain performance objective, thus achieving more consistent and reliable planning in unknown environments.

[0050] The aforementioned UAV adaptive motion planning method for different environments, targeting the UAV local planning problem based on visual depth observation, designs a robust dual-track mechanism. By optimizing the parallel processing model architecture design and training algorithm, it can mitigate the bias of the test distribution under multi-environment training, enabling the UAV to maintain reliable and safe planning behavior when the environmental geometry (different environments have different geometric features, corresponding to different state observation features) and observation features (corresponding to different depth observation features) change. It supports the generation of stable, continuous, and dynamically executable trajectories with significant safety margins in complex environments, thereby achieving cross-domain generalization and stable execution of short-term planning based on UAV vision. It improves stable cross-scene learning ability, cross-environment generalization ability, and cross-domain robustness, thereby enhancing the real-time autonomous flight performance of the UAV.

[0051] Specifically: In terms of model architecture design, the MoE architecture is adopted to enhance the model's ability to fuse different modal feature inputs from the environment (specifically: different expert networks are used to focus on the processing of visual features and state features in different feature streams, enabling them to learn the shared structure invariant to the environment and the differences specific to the environment; at the same time, the dynamic gating mechanism of sparse routing enables the model to adaptively select appropriate experts based on the local geometric features of different environments during the inference stage, thereby adaptively fusing visual features and state features to achieve robust cross-scene feature fusion).

[0052] In terms of training algorithm optimization, the DRO architecture is adopted, which transforms cross-environment policy learning into worst-case performance expectation optimization, thereby improving the model's performance in such challenging scenarios.

[0053] Furthermore, the proposed method was simulated and verified on a test set covering both in-distribution and out-of-distribution scenarios. Experiments show that in environments where the distribution changes significantly, the proposed method can generate more stable planning behavior, achieve more stable and reliable planning performance, and maintain higher trajectory quality in unknown environments. Moreover, even in scene domain images with poor pixel quality (below 20,000 pixels), reliable adaptive motion planning can be achieved, accelerating training speed, reducing latency and resource requirements, enabling real-time response, improving obstacle avoidance performance, and thus significantly enhancing cross-environment generalization ability.

[0054] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0055] In one specific embodiment, the planned behavior of the method of this application is visualized to visually demonstrate its obstacle avoidance decision-making and attention patterns in structurally complex environments.

[0056] like Figure 3 As shown, the starting point of the complete drone flight trajectory is a triangle symbol, and the ending point is a pentagram symbol. Three representative interactive areas are selected and marked as follows: , , Furthermore, by magnifying these three areas, which cover high-risk situations (such as traversing narrow passages, approaching obstacles, and dealing with continuous structural turns), a deeper examination of trajectory stability and safety margins under structural changes can be conducted.

[0057] like Figure 4 As shown, select the interactive area. The key moments are a, b, and c.

[0058] like Figure 5 As shown, select the interactive area. The key moments are d, e, and f.

[0059] like Figure 6 As shown, select the interactive area. The key moments: g, h, i.

[0060] Compare the original input depth image, pseudo-color depth visualization image, and Grad-CAM attention image at each key moment, such as... Figures 7 to 21 As shown (taking only the critical moments a, c, e, g, and h as examples).

[0061] from Figures 7 to 21 As can be seen, the pseudo-color depth visualization image more intuitively depicts the obstacle layout, where blue represents closer areas and red represents farther areas. The Grad-CAM attention image highlights the perceptual areas that have the greatest impact on the current decision, and its color mapping is the opposite of the pseudo-color depth visualization image, with warm colors (such as red and yellow) indicating higher relevance and cool colors (such as blue) indicating lower relevance.

[0062] The results show that the method in this application consistently focuses on the obstacle boundaries and their surrounding neighborhoods most relevant to collision avoidance, and this focus remains stable throughout the flight. This behavior helps explain the reliable planning performance observed in different scenarios and supports AERO's ability to generate continuous, dynamically executable trajectories with significant safety margins in complex environments. Furthermore, for scene domain images with poor pixel quality (… Figures 7 to 21 (The image uses a pixel size of 160×90). The method in this application can also achieve reliable adaptive motion planning.

[0063] To further verify the generalization ability, two randomly generated obstacle environments with significantly different structures were constructed in Gazebo for qualitative evaluation.

[0064] like Figure 22 The overall environmental layout map of the forest environment shown is as follows: Figure 23 A top-down view of the planning and execution trajectory from start to finish in the forest environment shown, as illustrated. Figure 24The overall environmental layout diagram of the columnar environment shown is as follows: Figure 25 The diagram shows a top-down view of the planning and execution trajectory from the starting point to the end point in a columnar environment. In both environments (including a layout environment resembling a dense forest and a mixed-scale environment containing columnar and blocky obstacles), the method of this application can autonomously identify passable paths and generate continuous, smooth detour routes amidst cluttered obstacles.

[0065] The above results qualitatively verify the adaptability of the method of this application to random obstacle distribution and significant changes in environmental geometry.

[0066] This application also provides an adaptive motion planning device for unmanned aerial vehicles (UAVs) in different environments, such as... Figure 26 As shown, in one embodiment, the UAV adaptive motion planning method for different environments includes: a first module 2601, a second module 2602, a third module 2603, a fourth module 2604, and a fifth module 2605, wherein: The first module 2601 is used to construct an initial overall planning network based on depth observation characteristics and state observation characteristics; The second module 2602 is used to construct the initial objective function based on the initial overall planning network; The third module 2603 is used to acquire scene domain images of different environments, sample them, and obtain the initial objective function for each sample; based on the initial objective function for each sample, the expected domain risk for each scene domain is obtained; a domain weight vector is introduced for the expected domain risk of each scene domain, and the weighted domain risk of all scene domains is calculated to obtain the worst-case domain objective function. The fourth module 2604 is used to update the domain weight vector according to the worst-case domain objective function in order to train the initial overall planning network until the preset conditions are met, and then obtain the optimized overall planning network. The fifth module, 2605, is used to acquire scene domain images of unknown environments, input them into the overall planning network for optimization, and output end-point state parameter constraints to construct the UAV's motion trajectory and achieve adaptive motion planning.

[0067] Specific limitations regarding the adaptive motion planning device for UAVs in different environments can be found in the limitations of the adaptive motion planning method for UAVs in different environments described above, and will not be repeated here. Each module in the aforementioned device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0068] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 27 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements an adaptive motion planning method for unmanned aerial vehicles (UAVs) adapted to different environments. The display screen can be an LCD screen or an e-ink display screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0069] Those skilled in the art will understand that Figure 27 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0070] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the method described above.

[0071] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0072] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0073] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0074] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended application documents.

Claims

1. An adaptive motion planning method for UAVs in different environments, characterized in that, include: Based on the characteristics of depth observation and state observation, an initial overall planning network is constructed; Based on the initial overall planning network, construct the initial objective function; Acquire scene domain images from different environments, sample them, and obtain the initial objective function for each sample; Based on the initial objective function for each sample, the expected domain risk for each scenario domain is obtained; For the expected domain risk of each scenario domain, a domain weight vector is introduced, and the weighted domain risk of all scenario domains is calculated to obtain the worst-case domain objective function; Using the worst-case domain objective function as a constraint, the domain weight vector is updated to train the initial global planning network until the preset conditions are met, thus obtaining the optimized global planning network. The system acquires scene domain images of unknown environments, inputs them into an optimized overall planning network, and outputs end-point state parameter constraints to construct the motion trajectory of the UAV and achieve adaptive motion planning. Based on the characteristics of depth observation and state observation, an initial overall planning network is constructed, including: Based on the depth observation features and state observation features, feature streams are extracted separately and connected along the channel dimension to obtain the fused input; The fused input is projected onto the latent representation, and multiple expert networks are introduced to obtain the output of each expert network. The latent representation is introduced into a router to obtain expert logical values, and a sparse gating method is used to establish an expert set. Normalization is performed within the expert set to obtain the gating weights. Based on the output of each expert network and the gating weights, the expert weighted sum is obtained; The expert weights are then mapped back to the original channel dimension to obtain the mapped input; Based on the mapping input, construct the initial global planning network; The fused input is projected onto the latent representation, and multiple expert networks are introduced to obtain the output of each expert network, including: In the formula, For potential representation, For feature fusion, To integrate input, For the first The output of an expert network, For the first A network of experts, The number of expert networks.

2. The adaptive motion planning method for UAVs in different environments according to claim 1, characterized in that, Based on the mapping input, construct the initial global planning network, including: In the formula, For the first The end state parameter constraints of the candidates, for The corresponding predicted cost, The number of candidate trajectories. To share the prediction head, For mapping input, For the initial overall planning network, for The depth observation characteristics at any given time for State observation characteristics at any given time For projection head, For hybrid expert modules, For feature fusion, For deep feature extractors, To spatialize state observations through coordinate transformation and grid alignment.

3. The adaptive motion planning method for UAVs in different environments according to claim 1 or 2, characterized in that, Based on the initial overall planning network, construct the initial objective function, including: Based on the initial global planning network, the end state parameter constraints are obtained, and the end state parameter constraints are parameterized in the world coordinate system to obtain the parameterized end state parameter constraints. The starting state is obtained, and combined with the parameterized ending state parameter constraints, an analytical polynomial generator is used to construct candidate trajectories. Construct differentiable cost labels based on candidate trajectories; Based on the differentiable cost label, trajectory loss and evaluation loss are established, and an initial objective function is constructed.

4. The adaptive motion planning method for UAVs in different environments according to claim 3, characterized in that, Based on the differentiable cost label, trajectory loss and evaluation loss are established, and an initial objective function is constructed, including: in, In the formula, Let the initial objective function be... For trajectory loss weighting coefficients, For trajectory loss, To evaluate the loss weighting coefficient, To assess the loss; The number of candidate trajectories. Differentiable cost label; To smooth the L1 loss function, for The corresponding predicted cost, For the first The end state parameter constraints of the candidates, To stop gradient propagation operator.

5. The adaptive motion planning method for UAVs in different environments according to claim 1 or 2, characterized in that, Based on the initial objective function for each sample, the expected domain risk for each scenario domain is obtained, including: In the formula, For the first Expected risks in each scenario domain. For the data distribution from the k-th scene Data distribution for samples Expectations The initial objective function for each sample, The number of scene domains.

6. The adaptive motion planning method for UAVs in different environments according to claim 5, characterized in that, For each scenario domain, a domain weight vector is introduced to represent the expected domain risk. The weighted domain risk for all scenario domains is then calculated to obtain the worst-case domain objective function, which includes: In the formula, For the model's learnable parameters, For the domain weight vector, Let K be the weight set of the scene domains. Weighted domain risk for all scenario domains For the first Domain weight vectors for each scene domain For the first Expected domain risks in each scenario domain.

7. The adaptive motion planning method for UAVs in different environments according to claim 1 or 2, characterized in that, Using the worst-case domain objective function as a constraint, the domain weight vector is updated to train the initial global programming network until preset conditions are met, resulting in an optimized global programming network, including: The worst-case domain objective function is used as the constraint for the initial global programming network to estimate the empirical risk, and the domain weight vector is updated using the exponential gradient ascent method. Based on the updated domain weight vector, the batch-level DRO loss is updated to train the initial global planning network until the preset conditions are met, resulting in an optimized global planning network.

8. An adaptive motion planning device for unmanned aerial vehicles (UAVs) in different environments, characterized in that: The adaptive motion planning method for unmanned aerial vehicles (UAVs) oriented towards different environments, as described in any one of claims 1 to 7, includes: The first module is used to construct an initial overall planning network based on the depth observation characteristics and state observation characteristics; The second module is used to construct the initial objective function based on the initial overall planning network; The third module is used to acquire scene domain images of different environments, sample them, and obtain the initial objective function for each sample; based on the initial objective function for each sample, the expected domain risk for each scene domain is obtained; a domain weight vector is introduced for the expected domain risk of each scene domain, and the weighted domain risk of all scene domains is calculated to obtain the worst-case domain objective function. The fourth module is used to update the domain weight vector according to the worst-case domain objective function in order to train the initial overall planning network until the preset conditions are met, and then obtain the optimized overall planning network. The fifth module is used to acquire scene domain images of unknown environments, input them into the overall planning network for optimization, and output end-point state parameter constraints to construct the UAV's motion trajectory and achieve adaptive motion planning.

Images