Forestry ground multi-robot cooperative full-coverage path planning method and device based on diffusion model

Abstract

The invention discloses a forestry ground multi-robot cooperative full-coverage path planning method and device based on a diffusion model, electronic equipment and a storage medium. The method comprises the following steps: constructing an enhanced environment model containing passing cost and reachability probability through multi-source sensor fusion; based on the model, generating a global path probability field with coherence and environmental adaptability by using a diffusion model; further realizing multi-robot self-adaptive task region division through a path flow field gradient; a local diffusion replanning mechanism is introduced for dynamic obstacles to carry out rapid path repair; and performing multi-objective optimization on the path in combination with a robot center Euclidean symbol distance field, and outputting an executable path meeting safety, smoothness and kinematics constraints. According to the method, through closed-loop design of perception, modeling, generation, division, re-planning and optimization, high-coverage, high-collaboration and high-robustness autonomous inspection of multiple robots in a complex forest region environment is realized, and reliable technical support is provided for smart forestry.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of autonomous navigation and intelligent path planning of mobile robots, and particularly relates to a forestry ground multi-robot cooperative full coverage path planning method and device based on a diffusion model, which is suitable for forestry inspection robots, under-forest ground mobile platforms, and forest resource investigation, forest road patrol, fire hazard monitoring, and pest and disease monitoring tasks in forest areas. BACKGROUND

[0002] With the increasing demand for intelligent forestry and ecological monitoring, forest inspection is transforming from manual patrol to unmanned mode with multi-robot cooperation. Manual inspection is low in efficiency, incomplete in coverage, and easily affected by fatigue and subjective factors, which is difficult to meet the requirements of efficiency, reliability and safety for large-scale and long-period forest operations.

[0003] The forest environment is dense in vegetation, significant in terrain undulation, and serious in shelter, with narrow and irregular traffic space, which makes traditional path planning face problems such as path interruption, insufficient coverage and repeated coverage in continuous inspection. Existing planning methods mainly include two categories: grid decomposition and random sampling. Grid methods (such as Boustrophedon, Spiral, etc.) are stable in regular environments, but are prone to high complexity or detailed coverage in areas with dense obstacles and variable terrain. Sampling methods (such as RRT, PRM) have exploratory ability, but are slow in convergence, and the paths are not smooth and coherent in narrow spaces, which is difficult to meet the requirements of high consistency and executability for large-scale inspection.

[0004] In terms of multi-robot cooperative planning, existing Voronoi division and topological division strategies are mostly based on the assumption of regular environment and stable communication. In the forest area with irregular structure and significant dynamic changes, these methods are prone to cause task block disconnection, uneven robot load or coverage redundancy, thereby affecting the cooperative efficiency and inspection quality.

[0005] In addition, traditional path planning mostly relies on fixed cost functions, which is difficult to comprehensively depict the obstacle distribution, accessibility, risk space and robot motion constraints in the forest area, and the generated path often has problems such as discontinuous curvature, insufficient smoothness and difficulty in execution, which is difficult to meet the requirements of coverage, safety and high precision for tasks such as forest fire monitoring and pest detection.

[0006] To solve the above bottleneck, the present application proposes a forestry multi-robot cooperative full coverage path planning method based on diffusion model enhancement. The method uses diffusion model to generate continuous and smooth global path probability field, and realizes adaptive task division between robots through flow field gradient, avoiding repeated coverage and path redundancy from the source. In the face of environmental changes, the local diffusion resampling mechanism can adjust the path in real time, ensuring the continuity and stability of the inspection. This method has strong environmental adaptability and autonomous decision-making ability, and is low in communication dependence, especially suitable for large-scale forest area scenes with complex structure and various obstacles, which can significantly improve the inspection efficiency and accuracy, and provide a reliable technical path for other multi-robot full coverage tasks. SUMMARY

[0007] In view of the problems of incomplete coverage, unreasonable multi-robot task division, high repetition rate and insufficient adaptability to complex dynamic environment in the existing forest inspection path planning, the present application proposes a forestry ground multi-robot cooperative full coverage path planning method, device, electronic equipment, storage medium and program product based on diffusion model enhancement. The present application integrates forest environment modeling, diffusion type global path field generation, adaptive cooperative partitioning and kinematic constraint optimization, and constructs an integrated planning system for complex forest ground surface, which is "perception-generation-collaboration-optimization-execution". Through diffusion model driven coverage path probability modeling and dynamic area reconstruction, the present application can generate continuous, smooth and safe cooperative inspection path with sufficient safety margin in large-scale forest area, effectively reduce missed inspection and repeated coverage, and significantly improve the multi-robot inspection efficiency and environmental adaptability.

[0008] I. Overall technical scheme and framework

[0009] The present application constructs a closed-loop path planning framework running through "environment modeling-path field generation-collaborative division-local re-planning-path optimization-execution control" around the forest ground inspection task, so that multiple ground robots can realize high coverage rate inspection, high robustness passing and cooperative operation under weak communication conditions in unstructured and dynamically changing forest environment. The overall process includes the following key steps:

[0010] (1) Forest environment modeling and reachability construction. The system completes high-precision environment mapping based on the fusion of multiple sensors such as laser radar, depth camera, IMU, odometer, GNSS / Beidou and millimeter wave radar. Through time synchronization, external parameter calibration and spatio-temporal fusion, fine-grained features such as slope, vegetation density, obstacle type and slippery risk are extracted, and an enhanced reachability model is constructed. This model can accurately depict the passing cost, especially in narrow forest paths or dense shrub areas, and improve the environment expression accuracy by high-density boundary sampling, providing a reliable basis for subsequent planning.

[0011] (2) Global path field generation based on diffusion model. The application generates a path probability field covering the entire area using a diffusion model. The system first diffuses the initial path distribution forward, which makes the path representation noisy and forms a smooth global distribution; then, combined with the terrain cost map, the environment structure is guided through the reverse denoising process, so that the path probability converges in the reachable area and presents a "natural avoidance" structure in the obstacle-dense or terrain-sudden-change area. This mechanism can adaptively adjust the denoising weight according to the environment, avoiding the problems of path breakage, insufficient coverage or topological distortion in irregular forest areas that occur in traditional grid and geometric decomposition methods, and forming a path probability field with continuity, coverage and consistent direction.

[0012] (3) Multi-robot adaptive cooperative partitioning based on path flow field. After obtaining the path probability field, the application constructs a path flow field according to the gradient of the probability distribution to extract the dominant flow direction, natural boundary and potential main line of the region. According to the characteristics of the flow field, the multi-robot task area is divided, so that each sub-area has good connectivity and naturally avoids impassable areas, avoiding the fragmentation, crossing obstacles or uneven load problems that may occur in traditional Voronoi or graph partitioning methods in forest areas. This data-driven partitioning method can maintain stable cooperation and clear task boundaries under weak or intermittent communication conditions.

[0013] (4) Local dynamic re-planning based on diffusion mechanism. To adapt to dynamic changes such as fallen trees, fallen branches and animal crossings, the application introduces a local diffusion re-planning mechanism. When the local environment changes suddenly, the system does not need to recalculate the global path, but only applies light noise to the local area and performs reverse diffusion, so that the path segment is quickly reconstructed without destroying the global structure. This mechanism supports millisecond-level response, so that the continuity of robot patrol and the stability of traffic can be maintained in dynamic scenarios.

[0014] (5) Multi-objective path optimization based on RC-ESDF. To ensure that the path meets the smoothness, safety and kinematic constraints of the robot at the execution level, the application constructs a multi-objective optimization model based on the robot-centric Euclidean signed distance field (Robo-centric ESDF, RC-ESDF). This model considers path length, curvature continuity, differential chassis speed- angular velocity constraints and whole-body level safety distance, and evaluates the minimum safety margin of the robot and obstacles such as tree trunks, rocks and shrubs through the safety distance field. The Prior-guided PSO algorithm is used to globally and locally optimize the diffusion-generated path, and to repair collision risks or curvature mutation segments, outputting a final patrol path with high executability and high safety.

[0015] (6) Comprehensive effect and technical advantage. Through the diffusion path generation, flow field driven cooperative division, local dynamic re-planning and path optimization based on safety distance field, the application forms a complete "perception-generation-cooperation-optimization-execution" closed-loop planning system. Compared with the prior art, the application can significantly improve the full coverage ability, cooperation efficiency and environmental adaptability of forest area inspection, and enhance the traffic stability and autonomous decision-making ability of the robot in irregular forest area and dynamic obstacle environment, and has good engineering landing performance and application potential in large-scale multi-robot cooperation scene.

[0016] II. Forest environment modeling and reachability construction

[0017] In order to cope with the unstructured, strong occlusion and terrain diversity of forest environment, the application constructs a multi-source sensor fusion modeling mechanism for forest scene. The system is configured with laser radar, depth camera, IMU, odometer, GNSS / Beidou and millimeter wave radar and other sensors, and the data of each sensor is mapped to a unified coordinate system through unified time synchronization and external parameter calibration. On this basis, a tightly coupled pose estimation method based on factor graph or extended Kalman filter (EKF) is used to jointly optimize the GNSS absolute position, IMU inertial measurement, odometer relative displacement and laser point cloud matching (ICP / NDT) results, and a stable and reliable continuous pose trajectory in the forest area is obtained, providing a high-precision reference for map construction.

[0018] In terms of map construction, the application combines grid map and high-resolution point cloud: two-dimensional grid map is used to identify passable and obstacle areas, and high-precision point cloud is used to extract terrain details. The system counts the height distribution in each grid unit and fits the local plane, calculates the slope, relief and ground roughness; based on the profile features of point cloud and the texture of depth camera, the rigid obstacles such as tree trunks and fallen trees and the passable vegetation are distinguished; the ground friction risk is estimated by the scattering characteristics of millimeter wave radar, the reflection intensity of depth map and the lateral acceleration change of IMU, and the additional cost is given to the slippage area such as mud and moss.

[0019] On this basis, the application converts the multi-dimensional environmental features f(x,y) of each grid [slope, roughness, vegetation density, wet and slippery risk, obstacle distance, passable weight] into the passable cost c(x,y) and the reachability probability p_reach(x,y) through the cost mapping function or the lightweight perception network. The slope and roughness punish steep or rugged terrain; the vegetation structure is used to distinguish the areas that can be lightly crossed and must be bypassed; the wet and slippery risk improves the cost evaluation of the slippery area, so that the reachability model reflects the geometric obstacles, surface physical properties and vegetation structure at the same time.

[0020] Considering the prevalence of narrow forest roads and bottleneck areas with dense obstacles in forest regions, this invention further introduces a high-density boundary sampling and resolution adaptive mechanism. The system first constructs a global grid at a medium resolution, then detects key bottlenecks based on boundary curvature, channel width, and obstacle density. For these areas, the grid is locally refined, increasing the resolution several times over, preventing misjudgment or fragmentation of passable space due to overly coarse grids. Ultimately, a high-quality accessibility environment map that accurately reflects the forest terrain, vegetation structure, and access conditions is obtained, providing reliable input for the subsequent diffusion model to generate a full-coverage path field.

[0021] III. Construction of Global Path Probability Field Based on Diffusion Model

[0022] After obtaining the accessibility environment map, this invention uses a diffusion model to generate a global path probability field covering the entire forest area, replacing the traditional grid coverage or geometric decomposition method that relies on manual rules, and realizing automated modeling of coverage and feasibility.

[0023] The system first describes the inspection task as: generating one (or more) continuous inspection paths that meet coverage requirements on a given access cost map c(x,y). The paths are represented as "access sequence fields" or "coverage marker fields" in the raster space, and Gaussian noise is gradually added during the forward diffusion process according to a preset noise schedule (such as linear or cosine schedule), resulting in a series of noisy path representations. Positive diffusion forms a smooth "coverage probability cloud" in passable areas, while the probability naturally decays in impassable areas (such as densely wooded areas, steep slopes, or bodies of water), thus establishing a preliminary feasibility boundary.

[0024] In the backdiffusion phase, this invention constructs a conditional denoising network. Using the environmental cost map c(x,y), reachability probability preach(x,y), and global task objectives (such as start and end points, and key inspection point sets) as conditional inputs, the noisy representation is restored to a structured path probability field P(x,y) through multi-step iterations. This process essentially "stretches" and fills in the path distribution within the passable area and automatically generates reasonable detours at obstacle boundaries, ensuring the final distribution possesses coherence, coverage, and directional consistency.

[0025] To better adapt to complex terrain, this invention incorporates a cost-guided diffusion mechanism during the diffusion process. In backsampling, the network adjusts the denoising direction based on the local cost gradient: enhancing probability decay in high-cost or high-risk areas and strengthening probability convergence in low-cost, flat areas, thus naturally skewing the path towards safe and feasible space. This mechanism directly integrates environmental constraints into the diffusion generation process, effectively avoiding problems such as path distortion and traversing infeasible regions that occur in traditional methods.

[0026] Meanwhile, to balance global coverage with local details, this invention employs a layered diffusion strategy: coarse-scale diffusion rapidly generates a backbone path distribution covering the entire area; fine-scale diffusion only supplements the bottleneck area or key inspection area with high resolution to refine the local probability structure. The resulting path probability field P(x,y) exhibits a clear "flow direction": globally, it presents an overall flow trend from the starting point to the forest and back to the exit, while locally, it forms flexible channels around obstacles such as tree trunks and rocks.

[0027] This continuous, manifold path probability field provides a natural and high-quality mathematical foundation for subsequent path flow field construction, multi-robot collaborative partitioning, and dynamic replanning.

[0028] IV. Multi-robot cooperative region division driven by path flow field

[0029] After obtaining the global path probability field P(x,y), this invention calculates the gradient field. The static probability distribution is transformed into a "path flow field" with directionality and traffic trends. The gradient direction represents the local optimal inspection trend, and the gradient magnitude reflects the intensity of traffic preference in the region. This invention utilizes this flow field structure to achieve adaptive region division for multi-robot collaborative inspection, ensuring that the sub-regions handled by each robot are both continuously reachable and conform to the real traffic topology.

[0030] Specifically, the system first performs streamline tracing in the path flow field: flowing along the gradient direction from the starting point or seed point to generate multiple path streamlines. Then, it clusters the streamlines based on spatial proximity and topological similarity, forming several internally connected but externally independent "basins." Each basin corresponds to a natural inspection sub-region, with consistent flow direction and continuous traffic within it, while different basins are typically separated by abrupt changes in flow direction or low-probability zones. Based on this, the system divides or merges basins according to the number of robots, task density, and single-robot load capacity, ensuring a balanced workload for each robot and a reasonable sub-region structure.

[0031] To prevent sub-regions from crossing infeasible areas or traversing large obstacles, this invention incorporates accessibility constraints into watershed division: regions with high path probability but insufficient accessibility are directly eliminated using a safety threshold; for regions connected only by narrow bottleneck channels, the decision to merge or split is based on the bottleneck width and robot passage capability, ensuring topological connectivity within sub-regions and that each sub-region can be independently covered by a single robot. Furthermore, a certain width of "cooperation zone" is reserved at the region boundaries to handle collaborative needs such as cross-regional intersections, yielding, and boundary patrol.

[0032] Considering that communication in forest areas is easily disrupted by trees and terrain, the area division in this invention can be completed offline. Before issuing tasks, the host computer generates a summary of the boundaries and flow directions of all sub-regions and sends it to the local control systems of each robot. During execution, each robot only needs to exchange progress and boundary conflict information at a low frequency to maintain cooperation, without continuously relying on high-bandwidth communication links, making it suitable for communication-constrained environments such as forest areas.

[0033] V. Local Dynamic Replanning Based on Diffusion Mechanism

[0034] The forest environment is highly dynamic; fallen trees, branches, temporary vehicles, and the entry of animals and personnel can all alter existing passage conditions. Therefore, this invention introduces a local dynamic replanning module based on a diffusion mechanism on top of global path planning, enabling rapid response to "local disturbances and local repairs." Local replanning is triggered when the system detects an overlap between the planned path and a new obstacle, a safety distance below a threshold, or a collision risk.

[0035] The system defines a local replanning window centered on the conflict point, extracts the path probability subfield Plocal(x,y) and the environmental cost subgraph clocal(x,y) from it, and marks the latest observed obstacle as a high-cost or impassable region. Unlike global diffusion, which requires multi-step sampling across the entire domain, this invention injects only a small amount of noise into Plocal(x,y) at the local stage to form a locally noisy representation. Then, under the constraint of the new environmental cost, a finite-step reverse diffusion is performed. Thanks to the generative properties of the diffusion model, this process can achieve natural detour around newly added obstacles by adjusting only local segments while keeping the surrounding path structure basically unchanged.

[0036] To reduce overhead, this invention employs an event-triggered and time-window suppression mechanism: replanning is initiated only when conflicts cannot be resolved through speed fine-tuning or slight yaw within a predicted future time domain; repeated triggering within the same area over a short period is suppressed to avoid oscillations. Simultaneously, the system maintains a "local topology cache," recording typical detour patterns for common obstacle changes (such as periodic vehicle stops or detour structures near fixed fallen logs). When similar situations recur, the cached path can be directly reused or used as initial values ​​for diffusion, thereby further shortening the replanning time.

[0037] Through the above design, the present invention can complete local route repair within milliseconds to hundreds of milliseconds, enabling the robot to continue inspection in sudden forest scenarios without frequent stops or large-scale retreats, thus greatly improving the stability and efficiency of operation in real forest environments.

[0038] VI. Multi-objective path optimization based on safe distance field

[0039] Building upon the initial path generated by diffusion and the locally replanned path, this invention introduces a multi-objective path optimization model based on the robot's central Euclidean symbolic distance field (RC-ESDF), further elevating the "feasible path" to an "engineerably executable path" that satisfies safety, smoothness, and dynamic constraints. Traditional planning only considers geometric reachability, neglecting the robot's shape and motion characteristics, often resulting in paths that appear feasible but are actually unsafe to traverse. This invention uses RC-ESDF to uniformly encode obstacle distance, robot size, and posture changes into continuously differentiable safety evaluation indicators.

[0040] Specifically, the system first binarizes the obstacle grid and calculates the distance from each grid to the nearest obstacle using Euclidean distance transformation. Then, it performs dilation calculations based on the robot's geometric model (such as a rectangle or ellipse) to map the obstacle distances to a feasible spatial distance field at the robot's center, obtaining the safe distance d(x,y). This distance directly reflects the minimum safety margin between the robot's entire body and the obstacles, and a safety cost term can be constructed, for example: Where φ(·) is a function that applies a steep penalty to insufficient safety distance.

[0041] After establishing the safe distance field, this invention constructs a comprehensive objective function:

[0042] J = α1L + α2C + α3D + α4S

[0043] Where L represents the total path length, which improves efficiency by penalizing unnecessary detours and redundant backtracking; C represents the degree of curvature change of the path, which ensures trajectory smoothness by penalizing abrupt changes in direction between consecutive path segments; D reflects the dynamic feasibility of the differential chassis, which, in combination with the upper limits of linear and angular velocities and acceleration constraints, penalizes situations where the switching between fast and slow speeds is too abrupt or the instantaneous steering is too sharp; S is the safety distance term, which evaluates the safety margin of each point on the path through RC-ESDF, and imposes a high penalty on path segments approaching obstacles to ensure that the robot has sufficient lateral and front-to-back safety distances in real forest areas.

[0044] In the optimization process, this invention employs a prior-guided particle swarm optimization (PSO) algorithm. The system uses the initially generated paths as search priors, encoding the coordinates of path points as particle positions, so that each particle corresponds to a candidate trajectory. The initial particle swarm is generated near this path through lightweight random perturbations, focusing the search on high-feasibility regions. During iteration, particles are continuously updated based on their historical best and globally best trajectories, and after each iteration, feasibility repair is performed using RC-ESDF, constraining and projecting path points that have exceeded limits or have insufficient safety margins, returning them to the safe region. Through this combination of prior guidance and safety constraints, the optimization process maintains effective exploration of the vicinity of high-quality trajectories while avoiding the generation of any unacceptably unsafe paths.

[0045] After multiple iterations, the particle swarm converges to an optimal trajectory that balances path length, smoothness, dynamic feasibility, and safety. The final generated path not only satisfies theoretical constraints but also exhibits good passability and execution stability in real forest scenarios such as dense forests, narrow forest roads, and areas with numerous obstacles, enabling it to reliably support multi-robot inspections over long periods.

[0046] VII. Apparatus and Implementation

[0047] Based on the same inventive concept, this invention provides a forestry ground multi-robot collaborative full-coverage path planning device based on a diffusion model, used to implement the main steps of the method of this invention on forest inspection robots or other autonomous ground mobile platforms. The device adopts a modular structure and includes the following functional modules:

[0048] (1) Multi-source sensing fusion and environment modeling module

[0049] This module receives and fuses data from multiple sources, including lidar, depth camera, IMU, odometer, GNSS / BeiDou, and millimeter-wave radar, to achieve data time synchronization, extrinsic parameter calibration, and coordinate unification. Tightly coupled pose estimation is achieved through factor graph optimization or extended Kalman filtering, outputting a robust continuous pose trajectory even in heavily obscured forest environments.

[0050] Building upon this foundation, the module performs rasterized modeling of point clouds and depth maps, extracting fine-grained surface features such as slope, roughness, vegetation density, obstacle contours, and slippery risk, and mapping them to access costs and accessibility probabilities to construct a high-fidelity accessibility environment model. For densely wooded and narrow forest road areas, it can adaptively increase local resolution to avoid misjudgments caused by insufficient resolution.

[0051] (2) Global path probability field generation module driven by diffusion model

[0052] This module uses an accessibility environment model to generate a path probability field covering the entire forest area through a diffusion model. The module represents the inspection task as a "path coverage field". First, it performs forward diffusion on the initial path representation to form a smooth global probability field. Then, it performs reverse denoising based on the access cost map, risk map, and inspection target to generate a path probability distribution with coherence and environmental constraints.

[0053] Through a cost-modulation denoising mechanism, the model can enhance path convergence in low-risk areas and naturally decay in high-risk areas, thus avoiding path breakage and unnatural detours common in traditional geometric scanning methods. The module supports coarse and fine-scale hierarchical diffusion, and performs high-resolution supplementation for key areas after generating the skeleton path in the entire domain.

[0054] (3) Path flow field driven multi-robot cooperative region division module

[0055] This module calculates the gradient of the path probability field, constructing a path flow field with directional and topological meaning. Based on streamline tracking and watershed clustering methods, it automatically identifies the main inspection flow direction, natural boundaries, and local trunk routes in the forest area, and completes the division of multi-robot inspection tasks accordingly.

[0056] The partitioning process comprehensively considers accessibility, bottleneck width, obstacle distribution, and load balancing, and adaptively adjusts the boundaries of sub-regions to avoid cross-obstacle partitioning and region fragmentation. Ultimately, it generates independent, connected, and load-balanced inspection sub-regions for each robot, which can execute independently under weak communication conditions.

[0057] (4) Path optimization and safety constraint module

[0058] This module is used for multi-objective optimization of the diffusion-generated path and the initial path after region partitioning. The module first constructs the robot center Euclidean symbolic distance field (RC-ESDF), and encodes the obstacle distance and robot geometry into a unified safety margin.

[0059] Based on RC-ESDF, an optimization function is established that integrates path length, curvature smoothness, dynamic feasibility, and safety distance. A prior-guided particle swarm optimization (PSO) algorithm is used to search for the optimal path in the feasible solution space. Constraint projection is applied to out-of-bounds or unsafe path points to ensure the path always meets safety requirements. The optimized path exhibits high smoothness, dynamic executability, and sufficient safety margin.

[0060] (5) Local dynamic replanning module based on diffusion mechanism

[0061] This module is used to handle new obstacles and sudden environmental changes during the inspection process. When insufficient safe distance or collision risk is detected, the module applies a small amount of noise to the path latent variables within a local window and performs finite-step conditional backdiffusion to quickly generate alternative path segments.

[0062] The module employs event triggering and time window suppression to avoid frequent replanning, and maintains a cache of typical detour paths to accelerate response in recurring scenarios. The path after partial replacement is then corrected using RC-ESDF to ensure the consistency and executability of the overall path.

[0063] (6) Data storage and management module

[0064] This module stores multi-source data such as environmental maps, accessibility models, path probability fields, region division results, replanning records, and optimization parameters. It supports task auditing, inspection playback, model training, and strategy optimization. The module can export paths and operation records and interface with the host computer system to achieve full traceability and supervision of the inspection process.

[0065] The above modules form a closed-loop system of "perception - modeling - path generation - region division - local replanning - path optimization - data management", which can run efficiently on embedded platforms and has real-time performance and engineering deployability.

[0066] This invention also supports an "edge-cloud collaborative" architecture: the cloud is responsible for model training, global policy updates, and large-scale data analysis; the edge performs online modeling, local diffusion replanning, and path tracking; and robots achieve state sharing and collaborative coverage through lightweight communication, improving task completion efficiency and stability.

[0067] The present invention also provides an electronic device for performing the above-described method, including a processor, a memory, and a communication interface, wherein the processor implements all or part of the steps of the present invention by running a program in the memory.

[0068] Meanwhile, the present invention also provides a non-transitory computer-readable storage medium and a computer program product, whose program instructions can be used to perform the entire process of forest environment modeling, diffusion path generation, regional division, local dynamic replanning and path optimization, etc., to realize the collaborative full-coverage inspection of multiple ground robots in complex forest environments.

[0069] VIII. Summary of Overall Innovative Points and Technical Effects of the Invention

[0070] This invention proposes a multi-robot collaborative full-coverage path planning method and device for forestry ground operations based on a diffusion model. Addressing the characteristics of complex terrain, irregular structures, diverse obstacles, and frequent dynamic changes in forest areas, it constructs an integrated intelligent planning system encompassing "multi-source perception—environmental modeling—diffusion-based path generation—flow field division—dynamic replanning—multi-objective optimization—closed-loop execution," achieving overall innovation from algorithm to system to collaboration level. Compared to traditional methods relying on geometric decomposition, rule coverage, or fixed cost functions, this invention achieves significant breakthroughs in path generation mechanisms, multi-robot collaborative strategies, and dynamic execution capabilities.

[0071] In terms of methodology, this invention is the first to apply a diffusion model to full-coverage planning in forest areas. By using a global path probability field to replace rule scanning, path generation can adapt to terrain undulations, vegetation structure, and obstacle distribution, thereby obtaining a coherent, natural, and executable global path structure. Based on this, this invention proposes a path flow field-driven region partitioning strategy, transforming the path probability distribution into a topologically meaningful flow direction structure. This effectively avoids the fragmentation, cross-obstacle partitioning, and uneven load distribution problems caused by traditional Voronoi or graph partitioning in forest areas, achieving high stability and rationality in multi-robot task partitioning. Simultaneously, the diffusion-based local dynamic replanning mechanism constructed in this invention can quickly generate alternative paths within local areas during environmental changes without global recalculation, significantly enhancing the system's real-time response capability and inspection continuity.

[0072] At the path optimization and execution level, this invention establishes a multi-objective path optimization model through the robot's central Euclidean symbolic distance field (RC-ESDF), enabling a unified trade-off between path length, curvature, dynamic feasibility, and safety margin. Combined with a priori guided particle swarm optimization strategy, it further improves optimization efficiency and final path quality, allowing the robot to maintain stable and safe passage performance in typical forest scenarios such as densely wooded areas and narrow forest roads.

[0073] At the system architecture level, the present invention adopts a modular design, which enables the device to be efficiently deployed on an embedded platform. Combined with an end-to-cloud collaborative mechanism and a lightweight multi-robot communication strategy, the robot group can maintain consistent task boundaries and stable collaborative behavior even under weak or even intermittent communication conditions, which significantly improves the overall efficiency and reliability of large-scale forest inspection tasks.

[0074] Overall, this invention achieves profound innovation in key technical aspects such as diffusion-based path generation, topology-consistent collaborative partitioning, dynamic replanning, and safety optimization control. It not only significantly improves the full coverage, robustness, security, and real-time performance of path planning but also demonstrates excellent engineering feasibility and cross-scenario adaptability. In addition to forestry inspection, this invention is also applicable to multi-robot coverage tasks such as agricultural inspection, park security, and geological inspection, providing a systematic and innovative solution for collaborative autonomous operations in large-scale, complex environments. Attached Figure Description

[0075] To more clearly illustrate the technical solutions of the present invention, the accompanying drawings used in the embodiments are now briefly described. It should be understood that these drawings are only used to illustrate some typical embodiments of the present invention, and those skilled in the art can obtain other forms of drawings or equivalent illustrations based on these drawings without creative effort.

[0076] Figure 1This is a schematic diagram of the overall process of the forestry ground multi-robot collaborative full-coverage path planning method based on the diffusion model provided in the embodiments of the present invention.

[0077] Figure 2 This is a schematic diagram of the module structure of the path planning device provided in an embodiment of the present invention.

[0078] Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0079] Before implementing the technical solution of this invention, relevant parties should, in accordance with applicable laws and regulations, fully inform users of any potentially involved personal information or data types, scope of use, purpose of use, and application scenarios, and obtain user authorization in accordance with the law. For example, when a user initiates an inspection task, uploads environmental data, or synchronizes work records, a prompt message can be used to inform the user of the data types required for this operation (such as location information, equipment identification, work logs, etc.), allowing the user to decide whether to authorize. The prompt can take the form of a pop-up window, a prompt box, etc., and the user can complete the authorization through interactive controls such as "agree" and "disagree". The above authorization process is only an example; other notification methods that comply with relevant laws and regulations are also applicable to this invention.

[0080] When collecting, transmitting, storing, and using environmental data, robot operation data, or information that may involve users, this invention shall comply with relevant requirements for data security and personal information protection, and take necessary technical and management measures to ensure that data processing is legal, proper, and necessary.

[0081] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the principles and implementation process of this invention are further explained below in conjunction with several exemplary embodiments. It should be understood that these embodiments are intended to help those skilled in the art understand this invention, and are not intended to limit the scope of protection; those skilled in the art can make equivalent modifications or substitutions based on these embodiments without any inventive effort.

[0082] The number of components shown in the accompanying drawings is for illustrative purposes only and does not constitute a limitation on the actual number; the reference numerals in the drawings are only used to distinguish different components and have no limiting meaning. Unless otherwise specified, the technical or scientific terms in this invention should be interpreted according to the common understanding of those skilled in the art. Terms such as "first" and "second" are only used to distinguish objects and do not indicate order or importance; "including" and "comprising" indicate that the listed contents may be all or part; "connected" and "linked" may refer to direct or indirect mechanical, structural, or signal connections. Positional terms such as "upper," "lower," "left," and "right" only indicate relative positional relationships and are adjusted according to the position of the objects, and do not constitute a limitation of this invention.

[0083] The following will describe in detail the core implementation process and structure of the present invention, "A Method and Device for Multi-Robot Collaborative Full-Coverage Path Planning in Forestry Based on a Diffusion Model," with reference to several representative embodiments and accompanying drawings.

[0084] In existing forest area ground inspection and multi-robot collaborative technologies, due to the complex terrain, irregular structure, dense obstacles, and frequent dynamic changes in forest areas, traditional coverage planning methods (such as regular grid scanning, geometric decomposition, or simple topology search) struggle to guarantee continuous coverage, collaborative stability, and path executability. This is especially true in real forest environments with large surface undulations, strong vegetation obstruction, and frequent dynamic obstacle changes, often resulting in discontinuous paths, unreasonable area division, insufficient dynamic response, and poor execution feasibility. Therefore, it is necessary to construct a comprehensive planning method and system that can balance full-area coverage, adaptive collaboration, and dynamic executability.

[0085] Based on the above technical background, this invention proposes a diffusion model-driven multi-robot collaborative full-coverage path planning method and device for forest scenarios. It constructs a complete collaborative planning chain through "multi-source perception fusion—enhanced accessibility modeling—diffusion-based path generation—path flow field partitioning—local dynamic replanning—multi-objective optimization." The core innovations include:

[0086] (1) Fine-grained accessibility modeling of forest areas driven by multi-source perception. By integrating sensors such as lidar, depth camera, IMU, GNSS / BeiDou, etc., a high-resolution accessibility model containing multi-dimensional access information such as slope, roughness, vegetation density, and slippery risk is constructed to provide physically consistent input for diffuse path generation.

[0087] (2) Construction of global path probability field based on diffusion model. The path generation is modeled as a diffusion-anti-diffusion conditional sampling process. Through global smooth representation and cost-guided denoising, a natural, coherent, and highly comprehensive path probability field is generated. It can automatically adapt to irregular terrain and obstacle distribution, and provide topological priors for collaborative planning.

[0088] (3) Direction guidance and initial path generation based on path probability field. By combining the dominant direction of the path probability field, environmental risk cost and safety distance (ESDF), a coherent and executable initial path is automatically formed, which maintains the path to advance rapidly in open areas and refines and adjusts it in areas with dense obstacles.

[0089] (4) Multi-robot collaborative region division based on path flow field. The path flow field is constructed by using the gradient of the path probability field, and inspection sub-regions with consistent topology, natural boundaries and balanced load are automatically formed to achieve stable collaboration under weak communication conditions.

[0090] (5) Multi-objective safety and executability optimization based on RC-ESDF. By integrating a multi-objective model that incorporates path length, curvature continuity, dynamic constraints and safety distance, the initial path is uniformly optimized to ensure that the final path has smoothness, dynamic executability and safety margin.

[0091] (6) Local dynamic replanning based on diffusion mechanism. For dynamic changes such as fallen trees, fallen branches or animals crossing, millisecond-level path repair is achieved through light noise disturbance and back diffusion of local conditions, without the need to recalculate the global path, thus ensuring the continuity of inspection tasks.

[0092] (7) Collaborative and consistent end-to-end planning closed loop. Through the unified constraints of diffusion-based global path prior, flow field division, local diffusion update and multi-objective optimization, multi-robot coverage planning with no overlap, no omission, executable and collaborative is achieved.

[0093] Having clarified the general principles of the present invention, the method flow, system structure, and device configuration will be further described below in conjunction with the accompanying drawings and embodiments. Without departing from the spirit of the present invention, those skilled in the art can make various equivalent substitutions and adjustments to the implementation methods.

[0094] refer to Figure 1 The forestry ground multi-robot collaborative full-coverage path planning method based on a diffusion model includes the following steps:

[0095] Step S101: Multi-sensor environmental data acquisition and surface feature extraction

[0096] During forest patrols, multiple ground robots utilize multi-source sensors, including LiDAR, depth cameras, IMUs, wheeled odometers, millimeter-wave radar, and GNSS / BeiDou positioning modules, to continuously and in real-time perceive the forest environment. The system first performs time synchronization, extrinsic calibration, statistical filtering, and dynamic outlier suppression on the sensor outputs to obtain a spatiotemporally consistent and noise-stable multimodal input sequence.

[0097] Subsequently, based on laser point cloud density gradient, ground / non-ground segmentation, vegetation penetration discrimination, and structured terrain analysis, fine-grained surface features such as slope distribution, vegetation density, root and shrub height hierarchy, surface roughness, access width, and slipperiness risk are extracted. These results constitute a high-dimensional environmental feature set including occupancy rate, slope field, vegetation distribution, and accessibility indicators, providing accurate and continuous basic inputs for accessibility modeling in step S102 and the construction of the diffuse path probability field in step S103.

[0098] Step S102: Environmental accessibility modeling and raster map construction

[0099] Based on the multi-source fusion data from step S101, the system first uses voxel grid filtering to control point cloud density and improve subsequent registration efficiency. Then, it employs improved ICP, NDT (Normal Distribution Transform), or multi-constraint graph optimization algorithms to achieve multi-frame point cloud registration, and utilizes tight-coupled state estimation between the IMU and wheeled odometer to reduce GNSS / BeiDou drift caused by forest obstruction.

[0100] After constructing a stable point cloud sequence, the system uses slope estimation, vegetation density estimation, and terrain profile analysis to generate an occupied raster map with accessibility as the core. Through point cloud density, texture gradient, and vegetation penetration analysis, the system further constructs a vegetation passability weight coefficient w_pass to distinguish between soft obstacles that can be squeezed through and hard obstacles that cannot be traversed.

[0101] Based on the aforementioned features, the system generates a comprehensive cost map C(x,y), which includes multi-dimensional terrain accessibility indicators such as occupancy rate, slope difficulty, vegetation resilience, passage width, slippery risk, and local roughness. Simultaneously, the system constructs an Euclidean symbolic distance field (ESDF) based on the occupancy grid and derives a safe distance field for the robot's center reference, used to evaluate the minimum obstacle safety margin at different locations.

[0102] The resulting multi-channel reachability model C(x,y) (where c(x,y) is the passage cost channel and preach(x,y) is the reachability probability channel) serves as the core input for the subsequent step S103 diffusion path probability field generation.

[0103] Step S103: Construction of global path probability field based on diffusion model

[0104] After constructing the accessibility map, the system introduces a diffusion model to generate a global path probability field P(x,y) covering the entire forest area. The diffusion model uses a mechanism of "positive noise perturbation + conditional reverse denoising" to uniformly map complex topology, dense obstacles, and changes in accessibility that are difficult to handle by traditional search methods into a continuous, smooth path distribution with environmental constraints.

[0105] Forward diffusion q(z) t |z t-1 The initial path latent variables are progressively noisy until a smooth global representation is achieved. Then, in the backdiffusion stage, the environmental cost model C(x,y) is used as the conditional input, and a conditional denoising network is used. θStepwise denoising is performed on the path's latent variables to guide them towards convergence to the passable region. The update format is as follows:

[0106]

[0107] Where, ∈ θ (z t C(x,y) is a conditional denoising network that integrates occupancy rate, slope, vegetation density, roughness, slipperiness risk, and traversability weights. It is used to predict and cancel the path latent variable z at time step t during backdiffusion. t The noise in the image makes the denoising result z t-1 Under the constraints of the environmental cost model C(x,y), it converges to the passable region. In the formula, z t z represents the path latent variable at time step t. t-1 The result is the denoised product after network correction; parameter α t The noise retention factor is the propagation noise retention factor. This is the cumulative retention factor, used to control the information retention ratio and the stability of reverse convergence.

[0108] The conditional denoising network ∈θ integrates multi-source terrain features such as occupancy rate, slope, vegetation density, surface roughness, slippery risk, and traversability weight. It can exert a stronger repulsive force on high-risk areas during backsampling and provide more path convergence degrees of freedom in open areas, thereby enabling the generated path representation to automatically tend towards areas with higher traversability potential and greater safety margin.

[0109] The path probability field P(x,y) generated after multiple back diffusion steps shows higher values ​​in forest passages, areas with high regional connectivity or with a safety margin, while it shows significantly lower values ​​in areas with dense fallen trees, steep slopes and slippery sections, thus achieving a continuous expression of the global path trend.

[0110] Step S104: Initial path generation based on path probability field and heuristic cost

[0111] After obtaining the global path probability field P(x,y) and the environmental cost graph C(x,y), the system combines these two to generate an initial feasible path that can be used for subsequent optimization. This process no longer relies on the traditional search pattern based on a fixed heuristic function, but directly utilizes the directional preferences given by the path probability field, and constructs a comprehensive cost function that incorporates environmental risks. For any node (x,y), the system defines the comprehensive cost as:

[0112] f(x,y)=g(x,y)+h(x,y),

[0113] Here, g(x,y) represents the cumulative cost from the starting point to the current node, and h(x,y) is the forward feasibility prediction from the node to the target based on the manifold direction of P(x,y) and the risk assessment of C(x,y). This is used to guide the search process to expand into safer areas with greater traversability potential.

[0114] To adapt to rapid changes in passage width, obstacle density, and local terrain undulations within forest areas, the system introduces an adaptive expansion step size mechanism, enabling automatic adjustment of node spacing according to changes in environmental risk. The system calculates the step size for each expansion using the following method:

[0115]

[0116] Where, d obs α represents the ESDF distance from the node to the nearest obstacle, used to quantify the safety margin at that location; α is the environmental sensitivity coefficient, used to adjust the step size as the node approaches the obstacle; k1 is the basic step size factor. The step size automatically shortens when the node approaches an obstacle, improving local obstacle avoidance accuracy; the step size automatically increases in open areas, improving search efficiency.

[0117] This mechanism ensures that the initial path has geometric coherence, environmental consistency, and basic kinematic feasibility, providing high-quality initial values ​​for the optimization phase (step S105).

[0118] The path generation results described above can be used for multi-robot task division and collaborative scheduling. The system can perform consistency maintenance at the region boundary to ensure full coverage and non-overlapping inspection results.

[0119] Step S105: Multi-objective path optimization and executable path generation based on RC-ESDF

[0120] After obtaining the initial global or local path, the system first performs a full-body safety margin assessment of the path using the robot's central Euclidean symbolic distance field (RC-ESDF). Based on this, a multi-objective joint optimization model is established to simultaneously constrain path safety, geometric continuity, and dynamic executability. This step does not generate a new path structure but optimizes the path of S104. The optimization objective can be expressed as:

[0121] J(P)=λ safe J safe +λ len J len +λ curv J curv +λ dyn J dyn

[0122] Among them, J safeThe minimum safe distance penalty between the path and obstacles such as tree trunks, rocks, fallen logs, and shrubs is calculated based on RC-ESDF to ensure sufficient obstacle avoidance margin is maintained throughout the inspection process; J len Characterize the total path length to avoid unnecessary detours or redundant bends; J curv Used to suppress abrupt angle changes and improve curvature continuity, resulting in a smoother trajectory; J dyn This constrains the kinematic conditions of the differential chassis, such as speed, angular velocity, and minimum turning radius, to ensure that the generated path can be strictly executed in a real robot.

[0123] The aforementioned weighting coefficient λ safe , λ len , λ curv , λ dyn >0 represents the weighting coefficients for security, length, smoothness, and dynamic executability, respectively, used to adjust the relative importance of each indicator in the optimization objective.

[0124] In forest patrol scenarios, safety is the highest priority, therefore, the following methods are typically adopted:

[0125] λ safe >>λ len ,λ curv ,λ dyn

[0126] Configure settings to ensure that optimization does not sacrifice obstacle safety margins by shortening paths or increasing smoothness, so that paths maintain reliable obstacle avoidance capabilities in scenarios such as dense forests, narrow forest roads, or high-risk slopes.

[0127] To efficiently solve this multi-objective optimization problem, the system employs a prior-guided particle swarm optimization algorithm (PSO), which utilizes the directional priors provided by the diffusion path probability field to shrink the search space, thereby improving convergence speed and solution feasibility. In each iteration, the optimizer performs constraint projection corrections on the potential collision points of particles based on RC-ESDF, ensuring that the optimization process always remains within the safe and feasible region, and gradually generates optimization paths that combine safety, smoothness, and dynamic consistency.

[0128] The optimized executable path is output in a structured manner using parameterization or hierarchical resolution, for stable tracking in multi-robot collaborative execution and weak communication environments.

[0129] Step S106: Local dynamic replanning based on diffusion perturbation

[0130] During actual inspection tasks performed by the robot, dynamic disturbances such as fallen trees, wild animals crossing the forest, and vehicles temporarily stopping frequently occur in the forest environment. These sudden events can cause local segments of the globally planned path to fail. When a local path failure is detected (such as the RC-ESDF safety distance falling below a threshold), in order to avoid high-cost global planning, the system uses a local diffusion replanning mechanism to quickly repair the affected path segments.

[0131] The system applies a lightweight random perturbation to the corresponding local path latent variables:

[0132]

[0133] Among them, z t Let represent the global path representation vector of the diffusion model at step t; σ is the local path latent variable to be added to the disturbance; σ is the noise intensity coefficient, used to adjust the range of the disturbance's influence; ∈ is zero-mean, unit-variance Gaussian random noise, used to pull the latent variable away from its original local convergence state.

[0134] After a local disturbance, the system performs conditional backpropagation on the latent variables of the path within the local window, causing them to rapidly converge to a new traversable path solution under the constraints of the environmental cost model C(x,y). This process operates only on a local space of a few meters in size, without disrupting the topology of the global path. Simultaneously, it can generate directly replaceable path fragments within milliseconds, ensuring that the ground robot can maintain continuous, safe, and uninterrupted inspection operations in dynamic forest environments.

[0135] For path segments that have undergone local replacement, the system splices them with the original global path and performs multi-objective optimization based on RC-ESDF again to uniformly correct safety margin, curvature continuity and chassis dynamics executability, thereby ensuring that the path after local reconstruction maintains consistent safety and executability at the global scale.

[0136] refer to Figure 2 The forestry ground multi-robot collaborative full-coverage path planning device based on a diffusion model of the present invention can be deployed on the inspection robot body, vehicle-mounted computing platform, or fixed edge node to execute the entire process of path generation, collaborative planning, and dynamic control. The device adopts a modular architecture, including a multi-source perception and environment modeling module 210, a path probability field generation module 220 driven by a diffusion model, a path flow field driven region division module 230, a local dynamic replanning module based on a diffusion mechanism 240, a path optimization and control execution module 250, and a data storage and management module 260. The modules exchange data and cooperate with each other through an internal communication interface.

[0137] The multi-source perception and environment modeling module 210 is used to receive and fuse data from multiple sources of sensors, including LiDAR, depth camera, IMU, odometry, GNSS / BeiDou, and millimeter-wave radar, during robot inspection. It performs time synchronization, extrinsic parameter calibration, and format consistency processing on the outputs of each sensor, and achieves robust pose estimation through factor graph optimization or extended Kalman filtering. Based on the fused pose, the module performs rasterization modeling on point cloud and depth information, extracting fine-grained surface features such as slope, roughness, vegetation density, slipperiness risk, and obstacle contours, and mapping them to accessibility probability and passage cost, forming a unified enhanced accessibility environment model. For narrow forest roads or areas with dense obstacles, the module can also adaptively increase the local raster resolution to provide more refined environmental input.

[0138] The diffusion model-driven path probability field generation module 220 is used to generate a global path probability field covering the forest patrol area based on the accessibility environment model. The module first performs multi-step forward diffusion on the initial path representation, transforming it from noise into a smooth global probability distribution. Then, it performs reverse denoising based on conditions such as accessibility cost, risk map, and patrol objectives, causing the path probability to converge in traversable areas and naturally decay in areas with dense tree trunks, fallen logs, or abrupt slope changes. Module 220 also dynamically adjusts the generation preferences for different terrain regions through a cost modulation denoising mechanism and supports hierarchical diffusion combining coarse and fine scales, thereby obtaining a path probability field that possesses both global structure and local detail.

[0139] The path flow field-driven region segmentation module 230 is used to transform the probability field into a directional path flow field and complete the sub-region segmentation for multi-robot collaborative inspection. The module calculates the spatial gradient of the probability distribution and assigns a forward direction and preference intensity to each grid cell; subsequently, it performs streamline tracing and clustering along the flow field to identify the dominant flow direction and natural boundaries in the forest area. Combining accessibility constraints, bottleneck channel width, and task load, module 230 automatically completes the segmentation of inspection areas with good connectivity, balanced load, and natural boundaries, and generates a flow direction description for each robot's corresponding region, which can be deployed locally for execution in one go under weak communication conditions.

[0140] The local dynamic replanning module 240, based on a diffusion mechanism, handles sudden environmental changes caused by fallen trees, branches, temporarily parked vehicles, or animal activity during inspections. When a conflict is detected between the planned path and a new obstacle or insufficient safety margin, the module constructs a local window around the conflict area, extracts the corresponding subfield from the global path probability field, applies a small amount of noise to the path representation within this window, and performs finite-step conditional backdiffusion to quickly generate alternative path segments. Module 240 avoids frequent replanning through event triggering and time window suppression logic, and can cache common detour patterns to further reduce replanning latency.

[0141] The path optimization and control execution module 250 is used to perform unified multi-objective optimization on the diffusion-generated path and replanned segments, and output low-level executable motion control commands. The module constructs a robot-centric Euclidean symbolic distance field (RC-ESDF) based on the environment map, evaluates the minimum safety margin of each path point, and establishes a comprehensive optimization objective that integrates path length, curvature smoothness, dynamic constraints, and safety distance. Through a priori-guided particle swarm optimization algorithm, the module iteratively updates path points under diffusion path direction constraints and safety distance constraints, performs constraint projection correction on path segments that exceed limits or have insufficient safety, and finally obtains an optimized path with smoothness, executability, and safety. It also generates corresponding speed control commands based on a differential kinematics model. When the execution deviation increases or the external environment changes abruptly, the module can trigger the local replanning module 240 to achieve closed-loop adjustment.

[0142] The data storage and management module 260 is used to uniformly store and manage multi-source data such as environmental maps, accessibility models, path probability fields, regional division results, replanning records, optimization parameters, and execution logs to support job auditing, quality assessment, model training, and cross-scenario migration and deployment. Module 260 can maintain the long-term accumulation of multi-round inspection data through a database or structured file system, and provides a data interface with a server or host computer system for long-term monitoring and inspection strategy optimization in forest areas.

[0143] It should be noted that, Figure 2 The modules 210, 220, 230, 240, 250, and 260 shown are merely exemplary structural divisions of the device of the present invention. In actual deployment, the modules can be merged, split, or distributed according to hardware conditions or task requirements. Without departing from the core concept of the present invention, any equivalent substitutions should be considered to fall within the protection scope of the present invention.

[0144] refer to Figure 3 The electronic device provided by this invention includes a processor 310, a memory 320, an input / output interface 330, a communication interface 340, and a bus 350. Each component interacts with data and instructions via the bus 350. The processor 310 can be a CPU, GPU, FPGA, DSP, ASIC, or a combination of multiple chips. It executes program instructions in the memory 320 to implement core logic such as environmental modeling, diffusion model inference, path probability field calculation, flow field generation, local diffusion replanning, RC-ESDF path optimization, and motion control instruction generation, thereby ensuring real-time path planning and control in complex forest environments.

[0145] The memory 320 may include ROM, RAM, Flash, solid-state drive, or other non-transitory storage media, for storing the operating system, device drivers, path planning program, diffusion model parameters, optimization model, environment map cache, and all software resources required for the path planning implementation of this invention. When this invention is implemented by software or firmware, its program code and model weights are stored in the memory 320 and loaded and executed by the processor 310.

[0146] The input / output interface 330 is used to connect local sensors and actuators, enabling environmental data input and robot motion output. Input devices may include LiDAR, depth cameras, IMUs, odometers, GNSS / BeiDou modules, millimeter-wave radar, buttons, or touch inputs; output devices may include displays, indicator lights, alarm modules, or actuator control interfaces. Through interface 330, the device can acquire sensor data in real time and send motion control commands to the robot chassis or controller.

[0147] The communication interface 340 is used to enable communication between the device and external systems. It can be wired via Ethernet, CAN, USB, or wireless via 5G, LTE, Wi-Fi, LoRa, satellite communication, etc. Through interface 340, the device can exchange environmental maps, area division information, execution status, or inspection logs with edge servers, other inspection robots, or cloud platforms to support collaborative inspections and remote management.

[0148] Bus 350 is used to connect processor 310, memory 320, interface 330, and interface 340, and undertakes internal data and instruction transmission tasks. Bus 350 may include address bus, data bus, and control bus to ensure smooth data flow and system stability when the device is running diffusion model inference and path planning tasks.

[0149] It should be noted that, Figure 3 The hardware structure shown is merely an exemplary implementation of the electronic device of this invention. Specific configurations can be adjusted according to the computing power, energy conditions, and deployment scenario of the robot platform. For example, an edge AI chip can be used to enhance the inference capability of the diffusion model, or some computing tasks can be migrated to the cloud for execution; alternatively, components can be trimmed and merged, retaining only the hardware required to implement the core steps of this invention. Equivalent substitutions made by those skilled in the art to the structural form without departing from the concept of this invention should fall within the protection scope of this invention.

[0150] The aforementioned electronic device can be used to execute the multi-robot full-coverage path planning method based on the diffusion model of any embodiment of the present invention, and its implementation effect is the same as that of the method embodiment, which will not be repeated here.

[0151] The present invention also provides a non-transitory computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, are used to implement any of the foregoing method embodiments. The storage medium may be a read-only memory, flash memory, solid-state drive, magnetic disk, optical disk, or other non-transitory medium capable of storing program code.

[0152] Meanwhile, this invention provides a computer program product comprising program instructions that, when executed by a processor, enable a computer or robot terminal to complete all or part of the steps of the path planning method of this invention. The software of this invention can be deployed in hardware, software, or a combination of both. The method flow of this invention can be implemented by storing the program code in the aforementioned storage medium and loading and running it by a processor.

[0153] It should be understood that the structure of the electronic device and program implementation is only an example and can be replaced or adjusted according to the computing power, power consumption limitations or deployment requirements of a specific platform. As long as it does not deviate from the core concept of this invention, it falls within the protection scope of this invention.

[0154] Furthermore, the method steps of this invention are presented in a specific order in the accompanying drawings, but are not limited to being executed in that order; the steps can be combined, adjusted, merged, or split as needed, and can also be executed in parallel. The flowcharts and block diagrams in the accompanying drawings are only schematic representations of possible system implementations, where each block can represent a program segment, logic code, or hardware functional unit, the function of which can be implemented by dedicated hardware or by software instructions executed on a general-purpose processor.

[0155] Those skilled in the art should understand that the above embodiments are only used to explain the technical principles of the present invention, and not to limit the scope of protection. Any modifications, substitutions, or equivalent changes within the spirit and principles of the present invention, including different software architectures, different storage media forms, or different hardware configurations, should be considered as included within the scope of protection of the present invention.

Claims

1. A forest ground multi-robot cooperative full coverage path planning method based on a diffusion model, characterized in that, The method comprises the following steps: Obtaining multi-source sensor data, constructing an enhanced accessibility model of the forest environment, wherein the accessibility model comprises a passing cost and an accessibility probability; Based on the accessibility model, a global path probability field covering the forest area is generated by using a diffusion model; Based on the path probability field, a path flow field is constructed, and multi-robot cooperative task area division is performed according to the path flow field; Based on the diffusion mechanism, local path conflicts caused by dynamic changes in the environment are re-planned; Based on the robot-centered Euclidean signed distance field, the initial path and the re-planned path are multi-objective optimized to generate an executable path that meets the safety, smoothness and dynamics constraints.

2. The method of claim 1, wherein, The construction of the enhanced accessibility model of the forest environment specifically comprises: Fusing multi-source sensor data of laser radar, depth camera, IMU, odometer, GNSS / Beidou and millimeter wave radar; Based on the fused data, the slope, roughness, vegetation density, slippery risk and obstacle profile fine-grained ground features are extracted; The ground features are mapped into a passing cost and an accessibility probability, and the local grid resolution is adaptively improved for narrow forest roads or obstacle-dense areas.

3. The method of claim 1, wherein, The generation of the global path probability field by using the diffusion model specifically comprises: Performing forward diffusion on the initial path representation to form a smooth global probability distribution; Performing reverse denoising based on the accessibility model to generate a path probability distribution with continuity and environmental constraints; Through a cost modulation mechanism, the path generation preference of different terrain areas is dynamically adjusted in the reverse denoising process.

4. The method of claim 1, wherein, The multi-robot cooperative task area division based on the path flow field specifically comprises: Calculating the gradient of the path probability field to construct a path flow field; Tracking and clustering the flow lines to identify the dominant flow direction and natural boundaries in the forest area; Combining the accessibility constraints, the bottleneck channel width and the task load, a good connectivity and load-balanced inspection sub-area is generated.

5. The method of claim 1, wherein, The local dynamic re-planning based on the diffusion mechanism specifically comprises: When detecting path conflicts or insufficient safety margins, a light noise is applied to the path representation in a local window; Performing limited-step reverse diffusion based on the local environmental cost to quickly generate an alternative path segment; Using an event-triggered and time-window suppression mechanism to avoid frequent re-planning.

6. The method of claim 1, wherein, The multi-objective path optimization based on the robot-centered Euclidean signed distance field specifically comprises: Constructing a comprehensive optimization objective function that integrates path length, curvature smoothness, dynamics feasibility and safety distance; Using a particle swarm optimization algorithm with prior guidance to iteratively optimize the diffusion path direction and safety distance constraints; Performing constraint projection correction on the path points that exceed the boundary or have insufficient safety margins.

7. A device for forestry ground multi-robot cooperative full coverage path planning based on a diffusion model, characterized in that, The method comprises: A multi-source perception and environment modeling module for constructing an enhanced accessibility model of the forest environment; A diffusion model-driven path probability field generation module for generating a global path probability field covering the forest area; A path flow field-driven area division module for realizing multi-robot cooperative task area division; A local dynamic re-planning module based on the diffusion mechanism for handling local path conflicts caused by dynamic changes in the environment; A path optimization and control execution module is configured to perform multi-objective optimization on the path and generate executable control instructions.

8. An electronic device, comprising: A computer program product comprising a memory, a processor and computer program stored on the memory and loadable on the processor, the processor implementing the method according to any one of claims 1 to 6 when executing the program.

9. A non-transitory computer-readable storage medium, comprising: The non-transitory computer readable storage medium stores computer instructions for causing a computer to perform the method according to any one of claims 1 to 6.

10. A computer program product, characterised in that, A computer program product comprising computer program instructions which, when loaded into a computer, cause the computer to perform the method according to any one of claims 1 to 6.

Images