A diffusion strategy-based obstacle avoidance trajectory generation method and system

Abstract

This application provides a method and system for generating obstacle avoidance trajectories based on a diffusion strategy, relating to the field of intelligent control technology. The method acquires the operating status information, task target information, and environmental status information of a mobile robot; determines the local planning area requiring local obstacle avoidance planning; identifies obstacle occupancy information within the local planning area; determines the location and local deformation of facility boundaries within the local planning area; and determines the effective passage boundary. Starting from multiple initial noise sequences, it performs multi-step inverse diffusion denoising iterations under the constraints of guiding condition features to generate multiple candidate obstacle avoidance trajectories. Candidate obstacle avoidance trajectories that conflict with obstacle occupancy information, exceed the effective passage boundary, or do not meet execution constraints are eliminated, and the target obstacle avoidance trajectory is determined from the remaining candidate trajectories. This method can improve the safety and passage efficiency of local obstacle avoidance trajectories in operating environments where the boundaries of facilities on both sides of the passage may deform.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and more specifically, to a method and system for generating obstacle avoidance trajectories based on a diffusion strategy. Background Technology

[0002] In factory-scale button mushroom cultivation facilities, mobile robots (such as four-way shuttles and AGVs) are responsible for tasks such as transporting mushroom beds, handling substrate, and conducting inspections. They need to frequently travel in the aisles between cultivation racks. Due to the limited width of the aisles, the possibility of multiple vehicles operating simultaneously, and the presence of dynamic or static obstacles such as workers, temporarily placed mushroom bed trays, and irrigation pipes within the aisles, generating local obstacle avoidance trajectories for the mobile robots within the aisles is a crucial step in ensuring operational safety and transportation efficiency.

[0003] Existing local obstacle avoidance methods for mobile robots, such as dynamic window methods, artificial potential field methods, and sampling-based local planning methods, typically treat the channel boundary as fixed static data read from a pre-built map. Based on this, they search for or optimize a single obstacle avoidance trajectory and then issue it for execution. These methods can meet basic usage requirements in scenarios where the channel boundary is stable and the obstacle conditions are relatively simple.

[0004] However, the above methods face two shortcomings in adaptability in industrialized cultivation facilities for button mushrooms.

[0005] On the one hand, button mushroom cultivation racks are typically multi-layered steel or aluminum alloy frames. Under the long-term influence of the weight of the mushroom beds, alternating temperature and humidity changes, and equipment vibrations, some sections may experience localized bending, tilting, or deformation towards the inside of the aisle. This deformation causes a deviation between the actual passable width of the aisle and the design width recorded in the pre-built map. When the mobile robot still plans its obstacle avoidance trajectory according to the aisle boundary corresponding to the design width, the generated trajectory may enter the area actually occupied by the deformed cultivation rack, causing the vehicle body or the carried mushroom beds to collide or scrape against the cultivation rack, resulting in equipment damage and potentially affecting the cultivated products of the mushroom beds.

[0006] On the other hand, in the aforementioned local areas where the passage width is limited and dynamic obstacles coexist, there is often more than one feasible obstacle avoidance path. For example, one can choose to avoid the obstacle to the left, avoid it to the right, or slow down and wait for the obstacle to pass before proceeding straight. This ambiguity makes traditional search- or optimization-based methods prone to the following situations: only one obstacle avoidance method is found and locked at a time, and when that method is no longer feasible, the search needs to be restarted; or the method repeatedly switches between two obstacle avoidance methods with similar costs, resulting in trajectory oscillations, which ultimately leads to decreased traffic efficiency or even forced stopping and waiting.

[0007] Therefore, how to generate local obstacle avoidance trajectories for mobile robots that balance safety and passage efficiency in working environments where the boundaries of facilities on both sides of the passage may deform is a technical problem that needs to be solved. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this application provides a method and system for generating obstacle avoidance trajectories based on a diffusion strategy.

[0009] Firstly, this application provides a method for generating obstacle avoidance trajectories based on a diffusion strategy, including:

[0010] Acquire the operating status information, task target information, and environmental status information of the mobile operation robot;

[0011] Based on the environmental status information, determine the local planning area for local obstacle avoidance planning; determine the obstacle occupancy information within the local planning area; determine the boundary position and local deformation of the button mushroom cultivation rack within the local planning area, and determine the effective passage boundary based on the boundary position and the local deformation.

[0012] The operating status information, the task target information, the obstacle occupancy information, and the valid passage boundary are encoded as guidance condition features;

[0013] Starting from multiple initial noise sequences, a multi-step inverse diffusion denoising iteration is performed under the constraints of the guiding condition features to generate multiple candidate obstacle avoidance trajectories.

[0014] Candidate obstacle avoidance trajectories that conflict with the obstacle occupancy information, exceed the effective passage boundary, or do not meet the execution constraints are eliminated, and the target obstacle avoidance trajectory is determined from the remaining candidate obstacle avoidance trajectories.

[0015] Optionally, determining the obstacle occupancy information within the local planning area includes:

[0016] Based on the current motion state of the dynamic obstacle, predict the predicted position of the dynamic obstacle at each moment in the planning time domain;

[0017] The obstacle occupancy information is determined based on the predicted location at each time point.

[0018] Optionally, determining the valid passage boundary includes:

[0019] Based on the design boundary of the button mushroom cultivation rack, determine the initial boundary of the channel;

[0020] The initial boundary of the channel is corrected along the inner side of the channel based on the local deformation to obtain the effective passage boundary.

[0021] Optionally, the multi-step inverse diffusion denoising iteration includes:

[0022] During the inverse diffusion denoising iteration, the trajectory samples are guided and updated based on the obstacle occupancy information, the effective passage boundary, and the execution constraints.

[0023] Optionally, determining the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories includes:

[0024] The remaining candidate obstacle avoidance trajectories are sorted according to the target positioning deviation and smoothness.

[0025] The candidate obstacle avoidance trajectory with the best ranking result is determined as the target obstacle avoidance trajectory.

[0026] Optionally, determining the local planning area for local obstacle avoidance planning based on the environmental state information includes:

[0027] The area where passage or arrival cannot be completed solely based on the predetermined driving trajectory under the current environmental conditions is identified as the local planning area.

[0028] Optional, also includes:

[0029] The mobile robot is controlled to execute the target obstacle avoidance trajectory in a rolling execution mode;

[0030] During execution, in response to the current execution trajectory no longer meeting the passage conditions or the deviation of the current position from the current execution trajectory meeting the replanning conditions, the target obstacle avoidance trajectory is regenerated based on the updated current state.

[0031] Optional, also includes:

[0032] Determine the region type of the local planning area in the mushroom growth area, harvesting area, or buffer passage, and determine the regional temperature and humidity transition parameters of the local planning area along the candidate passage direction;

[0033] Based on the cross-regional path, running speed, and regional temperature and humidity transition parameters corresponding to the candidate obstacle avoidance trajectory, the regional environmental disturbance cost is determined; the regional environmental disturbance cost is used to characterize the degree of influence of the candidate obstacle avoidance trajectory on the stability of the mushroom growth environment and the smoothness of the cross-regional environmental transition.

[0034] When determining the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories, the regional environmental disturbance cost is used as one of the sorting criteria.

[0035] Optionally, obtaining the effective passage boundary includes:

[0036] Determine the passage envelope height range of the mobile operation robot under the carrying conditions;

[0037] Within the height range of the passage envelope, the deformed intrusion sections of the mushroom cultivation rack that intrude into the inner side of the channel relative to the initial boundary of the channel, and the intrusion amount corresponding to each deformed intrusion section are determined;

[0038] Based on the intrusion amount corresponding to each deformed intrusion segment, the initial boundary of the channel corresponding to the deformed intrusion segment is shrunken and corrected, and then spliced ​​with the initial boundary of the channel corresponding to the non-deformed intrusion segment to obtain the effective passage boundary.

[0039] Secondly, this application provides an obstacle avoidance trajectory generation system based on a diffusion strategy, comprising:

[0040] The data acquisition module is used to acquire the mobile robot's operating status information, task target information, and environmental status information.

[0041] The first processing module is used to determine the local planning area for local obstacle avoidance planning based on the environmental state information; determine the obstacle occupancy information within the local planning area; determine the boundary position and local deformation of the button mushroom cultivation rack within the local planning area; and determine the effective passage boundary based on the boundary position and the local deformation.

[0042] The encoding module is used to encode the running status information, the operation target information, the obstacle occupancy information, and the effective passage boundary into guidance condition features;

[0043] The second processing module is used to perform multi-step inverse diffusion denoising iteration based on multiple initial noise sequences and under the constraints of the guiding condition features to generate multiple candidate obstacle avoidance trajectories.

[0044] The selection module is used to eliminate candidate obstacle avoidance trajectories that conflict with the obstacle occupancy information, exceed the effective passage boundary, or do not meet the execution constraints, and to determine the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories.

[0045] Compared with existing technologies, this application does not rely solely on obstacle information in the environment or the design of passage boundaries for local trajectory planning. Instead, it further considers the actual location and local deformation of the facility boundaries within the local planning area, and determines the effective passage boundary based on this. In this way, the passage range on which the trajectory generation process is based no longer remains in the ideal design state, but can more closely reflect the actual passability conditions in the working environment, thereby solving the problem of nominal passability but high actual execution risk caused by planning based on static boundaries.

[0046] Simultaneously, this application uses operational status information, task target information, obstacle occupancy information, and effective passage boundaries as guiding conditions, introducing a multi-step inverse diffusion denoising iterative process to generate multiple candidate obstacle avoidance trajectories. Therefore, it can not only generate multiple candidate passage methods in complex scenarios with dynamic obstacles and boundary deformations, but also remove trajectories that conflict with obstacle occupancy, exceed the effective passage boundary, or do not meet execution constraints through subsequent screening mechanisms, thus obtaining a target obstacle avoidance trajectory more suitable for actual execution. Based on the above technical solution, this application can improve the adaptability of local trajectories to changes in real-world environmental boundaries, and enhance passage safety, trajectory executability, and overall passage efficiency in narrow passages or complex operational environments. Attached Figure Description

[0047] Figure 1 A flowchart illustrating an obstacle avoidance trajectory generation method based on a diffusion strategy, provided for an embodiment of this application;

[0048] Figure 2 A flowchart illustrating a method for determining obstacle occupancy information provided in this application embodiment;

[0049] Figure 3 A flowchart illustrating a method for determining a target obstacle avoidance trajectory provided in an embodiment of this application;

[0050] Figure 4 This is a schematic diagram of an obstacle avoidance trajectory generation system based on a diffusion strategy, provided in an embodiment of this application. Detailed Implementation

[0051] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0052] See Figure 1 The diagram shows a flowchart of an obstacle avoidance trajectory generation method based on a diffusion strategy provided in an embodiment of this application, including steps S101 to S105, wherein:

[0053] S101: Obtain the operating status information, task target information, and environmental status information of the mobile robot;

[0054] S102: Based on the environmental status information, determine the local planning area for local obstacle avoidance planning; determine the obstacle occupancy information within the local planning area; determine the boundary position and local deformation of the button mushroom cultivation rack within the local planning area, and determine the effective passage boundary based on the boundary position and the local deformation.

[0055] S103: Encode the running status information, the operation target information, the obstacle occupancy information, and the effective passage boundary into guidance condition features;

[0056] S104: Starting from multiple initial noise sequences, perform multi-step inverse diffusion denoising iteration under the constraints of the guiding condition features to generate multiple candidate obstacle avoidance trajectories;

[0057] S105: Eliminate candidate obstacle avoidance trajectories that conflict with the obstacle occupancy information, exceed the effective passage boundary, or do not meet the execution constraints, and determine the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories.

[0058] The mobile operation robot of this application can be a transport robot that performs transfer, connection, alignment or loading and unloading tasks in the industrialized cultivation environment of button mushrooms. Its specific chassis structure, drive method, sensor type and controller implementation method can be selected according to the actual application, and those skilled in the art can use conventional technical means to implement it.

[0059] In one embodiment, the execution can be performed by a controller mounted on the mobile robot body, or by an edge control node or a higher-level scheduling control unit communicatively connected to the mobile robot. The controller may include a processor, a memory, and program instructions stored in the memory. When the program instructions are invoked by the processor, steps S101 to S105 are executed. The acquisition and updating of environmental status information, operational status information, and operational target information can be obtained through lidar, depth cameras, millimeter-wave radar, encoders, inertial measurement units, odometry, position markers, area controllers, or any combination thereof. These can all be implemented using conventional sensing and state estimation techniques in the art, and this application will not elaborate on conventional hardware connection methods and basic positioning principles.

[0060] Regarding S101:

[0061] In one embodiment, the operational status information may include at least one of the following: the current position, current orientation, current linear velocity, current angular velocity, and current carrying status of the mobile robot. The current position and current orientation can be obtained by the robot's localization module based on an environmental map and sensor observations using conventional localization methods. The current linear velocity and current angular velocity can be obtained from feedback by a wheel speed encoder, inertial measurement unit, or drive controller. The carrying status can characterize whether the robot is in an empty, half-loaded, or fully loaded state, or whether the robot is currently carrying a mushroom cultivation tray, cultivation bed, or other transport carrier. The operational status information is obtained because, under the same channel conditions, the braking capacity, sway amplitude, and acceptable passage margin may differ depending on the speed and carrying status. Without considering this information, it is easy to encounter situations where candidate trajectories are geometrically passable but difficult to stably track in actual execution.

[0062] In one embodiment, the task target information is used to characterize the target task location or target passage task served by the local obstacle avoidance planning. It may include at least one of the following: target location, target entry direction, target parking posture, target passage direction, or target alignment area. For example, in a mushroom cultivation workshop, when a mobile robot needs to transfer a vehicle carrying cultivation materials from a buffer channel to the vicinity of a designated picking station, the task target information can not only characterize the location of the target station, but also characterize the preferred entry direction adopted by the robot when approaching the station, so that the subsequently generated local trajectory is more conducive to completing the task.

[0063] In one embodiment, environmental state information is used to characterize the distribution of static facilities, dynamic obstacles, and changes in the passable area in the local environment. It may include at least one of the following: the local environmental point cloud at the current moment, obstacle detection results, facility boundary observation results, and area occupancy status. For the industrialized cultivation scenario of button mushrooms, environmental state information typically includes not only dynamic obstacle information such as personnel, other transport robots, and temporary parking vehicles, but also static facility information such as button mushroom cultivation racks, passageway edges, and supporting components. This application does not limit the specific carrier form of the environmental state information; it can be raw sensing data, or a raster occupancy map, boundary contour set, obstacle list, or area state table converted using conventional preprocessing methods in the art.

[0064] Regarding S102 above:

[0065] In one embodiment, the local planning area refers to the area where, under the current environmental conditions, the local obstacle avoidance planning module needs to directly generate a passage trajectory. This area can be a local passageway in front of the robot's current direction of travel, a local alignment area when approaching the target workstation, or a local passageway that needs to be re-circuited due to the intervention of dynamic obstacles. The local planning area can be determined using conventional methods in the art. For example, using the robot's current pose as a reference, a local planning window can be formed by intercepting a preset length and a preset width along the current direction of travel and the target direction; or, the local planning range can be formed based on the local passable area between the target position and the robot's current position. The scale of the local planning area can be comprehensively set according to the robot size, passage width, braking distance, and effective perception distance. For example, in the narrow passage scenario of a button mushroom cultivation workshop, the forward length of the local planning area can be set to 1.5m to 6m, and the lateral width can be set to the passage design width plus a safety extension. For example, when the robot operates at a low speed, the passage is short, and precise alignment is required, the forward length can be 2.0m; when the robot needs to make detour decisions in advance, the forward length can be 4.0m. These values ​​are not isolated settings, but correspond to the planning requirements of close-range precise alignment scenarios and medium-to-short-distance local detour scenarios, respectively.

[0066] In one embodiment, determining obstacle occupancy information within a local planning area may include identifying dynamic obstacles within the local planning area based on environmental state information, further estimating the predicted positions of the dynamic obstacles at each moment in the planning time domain, and then forming obstacle occupancy information based on the predicted positions at each moment. Obstacle occupancy information refers to information used to characterize the occupancy status of dynamic or static obstacles on the passage space of the local planning area within the planning time domain. It may be a time-by-time set of occupancy areas, a time-marked occupancy grid, an obstacle envelope sequence, or other structured information that can be used for trajectory collision judgment. For the predicted positions of dynamic obstacles, conventional prediction methods in the art can be used, such as constant velocity extrapolation, constant acceleration extrapolation, smooth prediction based on trajectory history, or short-term prediction based on behavioral patterns. The planning time domain can be set according to the robot's running speed, channel length, and control refresh cycle. For example, the planning time domain can be set to 1s to 5s; in scenarios requiring rapid response to local disturbances, it can be 1.5s to 2.5s; and in scenarios requiring dynamic avoidance of intersecting channels ahead, it can be 3s to 4s. The time interval between walks can be set from 0.05s to 0.2s, for example, 0.1s can be chosen to balance prediction accuracy and computational load.

[0067] In one embodiment, the boundary position of the button mushroom cultivation rack is used to characterize the current actual edge position of the cultivation rack, supporting column, frame member, or load-bearing component adjacent to the aisle within the local planning area. The boundary position can be obtained by edge extraction, contour fitting, planar projection, or elevation projection of environmental perception data, or it can be determined by registering the observed boundary with the design model. All of the above processing methods can be implemented using conventional geometric processing methods in the art. Unlike rigid standard shelves in general warehousing scenarios, button mushroom cultivation racks may experience local bulging, edge offset, local shelf sagging, or slight outward expansion of components under long-term use, load changes, local stress, or environmental disturbances. Although such changes may not significantly alter the overall structural layout, they may compress local passage space, thereby affecting the actual passage capacity of robots in narrow aisles.

[0068] In one embodiment, to improve the detection accuracy of local boundary deformation of the mushroom cultivation rack, a line lidar, an industrial structured light depth camera, or a combination thereof can preferably be used to scan the side of the passage. For example, within a detection range of 0.5m to 1.5m between the robot and the boundary of the cultivation rack, the repeatability of the line lidar or industrial structured light depth camera can reach ±2mm to ±5mm; after combining multi-frame averaging, boundary fitting, or outlier removal, the detection requirements for local deformations exceeding 5mm can be met. This provides a practically feasible measurement basis for subsequent calculation of local deformation and correction of the effective passage boundary.

[0069] In one embodiment, the local deformation is used to characterize the degree of local offset of the mushroom cultivation rack relative to the design boundary or reference boundary towards the inside or outside of the channel. The local deformation can be obtained by comparing the current observed boundary with the design boundary, or by comparing the current boundary position with the reference boundary position under historical stable operating conditions.

[0070] Local deformation is a geometric quantity reflecting the offset of the facility boundary relative to the reference boundary within a local section. It can be expressed as normal offset, boundary encroachment, edge bulging, or local deformation amplitude. In one embodiment, the cultivation rack boundary can be discretely sampled along the channel extension direction to form multiple boundary sampling points. The normal offset value of each sampling point relative to the reference boundary is then calculated as the local deformation at the corresponding location. The sampling interval can be set according to the robot body length, control resolution, and channel boundary change frequency. For example, it can be set to 10mm to 80mm; 50mm can be used in scenarios with relatively gentle boundary changes, and 20mm can be used in scenarios where there is a significant risk of bulging or misalignment at the local boundary.

[0071] In one implementation, determining the effective passage boundary based on boundary location and local deformation means that instead of directly using the designed passage boundary as the boundary condition for local obstacle avoidance planning, the passage boundary within the local planning area is modified based on the currently observed actual boundary location and local deformation to obtain an effective passage boundary that more closely reflects the actual passable range. Here, the effective passage boundary refers to the real boundary condition used to constrain candidate trajectory generation and screening; it reflects the passage edge that the robot can safely pass through under the current facility boundary conditions. Compared to the designed boundary, the effective passage boundary focuses more on actual passability than on the design drawing state. By introducing the effective passage boundary, it is possible to avoid generating trajectories that are theoretically passable but actually have a high risk of execution, even when the theoretical passage width meets the passability conditions but the actual local boundary encroachment leads to insufficient passability margin.

[0072] In one embodiment, when local deformation is used to correct the effective passage boundary, the initial boundary of the passage can be determined first, and then the local segment can be corrected inward according to the degree of boundary encroachment. The effective threshold for determining local deformation can be set in conjunction with the robot's lateral control error, the vehicle's envelope width, and the safe passage margin. For example, the threshold for local deformation can be set to 5mm to 40mm. For instance, when the safety margin between the robot body and the passage boundary is small and high-frequency passage is required, the threshold can be set to 8mm; when the passage is wide and the control precision is high, the threshold can be set to 20mm. These threshold settings correspond to the boundary correction requirements in scenarios with narrow safety margins and relatively wide safety margins, respectively.

[0073] Regarding S103:

[0074] In one implementation, guiding condition features are used as constraining inputs when the diffusion strategy model performs multi-step inverse diffusion denoising iterations. That is, the diffusion strategy model does not arbitrarily generate trajectories from random noise under unconstrained conditions, but rather gradually converges to generate candidate trajectories adapted to the current operational scenario, under the joint constraints of the operating state, work objective, obstacle occupancy, and effective passage boundaries. Guiding condition features refer to conditional information expressions that can impose constraints or guide the generation direction and distribution of candidate trajectories.

[0075] In one embodiment, obstacle occupancy information can be discretized into multiple local occupancy grids according to the planning time domain. The effective passage boundary is represented as a boundary constraint band in each corresponding frame. The running status information and task target information are organized into a low-dimensional condition vector. Then, the multiple local occupancy grids, boundary constraint bands, and low-dimensional condition vectors are fused to form a unified guidance condition feature. Specifically, the multiple local occupancy grids characterize the change in obstacle occupancy over time within the planning time domain, the boundary constraint bands characterize the actual passable range within the local planning area, and the low-dimensional condition vectors characterize the robot's current running status and target arrival requirements. Through this organization, the diffusion strategy model can simultaneously perceive changes in obstacle occupancy, effective passage boundary restrictions, and target task requirements in each denoising iteration step.

[0076] In one embodiment, the spatial resolution of the local occupancy grid can be set according to the channel width, the robot's lateral control accuracy, and the perception resolution. For example, the grid cell side length can be set to 10mm–80mm; in scenarios with narrow channels and sensitive boundary deformation, 20mm or 30mm can be used; when the local planning area is large and the overall passage trend is more important, 50mm can also be used. The number of time frames can be determined according to the planning time domain length and time interval step size. For example, when the planning time domain is 2s and the time interval step size is 0.1s, 20 occupancy grid frames can be formed. The width of the boundary constraint band can be set according to the minimum safe passage margin. For example, it can be set to 5mm–30mm to explicitly represent the inaccessible area near the effective passage boundary as a constraint band. The above parameters correspond to different feature organization requirements for fine narrow channel local planning scenarios and relatively wide area planning scenarios, respectively.

[0077] In one embodiment, the low-dimensional conditional vector may include at least one of the following: current position, current orientation, current velocity, target position, and target orientation. If the current task is sensitive to the approach direction, a representation of the target approach direction or the target docking posture may be further added. This application does not limit the specific network structure of feature fusion; any structure that can form a unified conditional input that simultaneously represents obstacle occupancy, real boundary constraints, and target task requirements is acceptable.

[0078] Regarding S104:

[0079] In one implementation, multiple initial noise sequences are used to characterize multiple trajectory generation starting points, enabling the diffusion strategy model to generate various candidate passage methods under the same scene constraints. For local narrow-channel obstacle avoidance problems, there is often no single reasonable path. For example, within the same local planning area, the robot can choose to detour slightly to the left, detour slightly to the right, slow down and wait before passing, or fine-tune its posture before approaching the target position. Therefore, using multiple initial noise sequences as starting points helps generate multiple differentiated candidate trajectories under the guidance condition feature constraints, rather than obtaining only a single trajectory solution.

[0080] In one embodiment, the multi-step inverse diffusion denoising iteration can be performed as follows: First, construct a corresponding initial trajectory sample for each initial noise sequence. The initial trajectory sample can be represented as a sequence of position points, attitude points, or control variables corresponding to multiple discrete moments in the planning time domain. Then, in each denoising iteration step, input the current trajectory sample and the guiding condition features into the diffusion strategy model to obtain the trajectory noise estimation result or trajectory correction result corresponding to the current iteration step. Then, update the current trajectory sample according to the trajectory noise estimation result or trajectory correction result to obtain the trajectory sample for the next iteration step. Repeat this process for multiple iteration steps until a candidate obstacle avoidance trajectory that satisfies the current scene constraints is obtained. In other words, the inverse diffusion denoising process does not directly output the complete trajectory at once, but gradually updates the trajectory sample to gradually converge to a trajectory distribution that meets the requirements of local obstacle occupancy, effective passage boundaries, and operational objectives.

[0081] In one implementation, after each update, the trajectory samples can be further guided for feasibility by incorporating obstacle occupancy information, effective passage boundaries, and execution constraints. Specifically, the updated trajectory samples can be checked to see if they have entered obstacle occupancy areas, exceeded effective passage boundaries, or exhibited excessive directional changes or state changes exceeding current control capabilities within the planning time domain. If any of these conditions exist, constraint corrections can be applied to the corresponding trajectory samples, causing the trajectory samples in subsequent iterations to converge towards more executable directions. In this way, constraint information is not only used for final screening but also participates in the formation of candidate trajectories, thereby increasing the proportion of effective candidate trajectories generated in the final step.

[0082] In one implementation, constraint correction can be achieved using a projection-based sample correction method. Specifically, after updating the current trajectory sample in each denoising iteration step, it is possible to check point by point whether the trajectory points corresponding to each discrete time in the planning time domain have entered the obstacle-occupying area or crossed the effective passage boundary. If a trajectory point enters the obstacle-occupying area, the shortest intrusion distance of the trajectory point relative to the outer boundary of the obstacle-occupying area can be calculated, and the trajectory point can be projected to the outside of the obstacle-occupying area along the shortest outward normal direction; if a trajectory point crosses the effective passage boundary, the cross-boundary distance of the trajectory point relative to the effective passage boundary can be calculated, and the trajectory point can be projected back to the inside of the effective passage boundary along the inner normal direction of the boundary.

[0083] In one embodiment, to avoid excessive local trajectory distortion caused by direct projection of a single trajectory point, continuity correction can be applied to the trajectory point after projection correction and its adjacent trajectory points. For example, within a range of 1 to 3 trajectory points before and after the current trajectory point, the trajectory segment can be adjusted for continuity using weighted averaging, linear interpolation, or local spline smoothing to reduce the impact of local abrupt changes on subsequent denoising iterations and trajectory smoothness. In this way, after each update, the trajectory sample not only returns to the area outside the obstacle-occupied region or within the effective passage boundary but also maintains continuity between adjacent trajectory points.

[0084] In one embodiment, for trajectory samples that violate execution constraints, corresponding corrections can be performed based on the type of violation. For example, when the rate of change of direction between adjacent trajectory points exceeds the maximum rate of change of steering, the trajectory sample can be corrected by reducing the amount of change of direction of the adjacent trajectory segment; when the velocity change between adjacent time points exceeds the allowable range, the displacement increment of the corresponding trajectory segment can also be reduced.

[0085] In one implementation, the number of initial noise sequences and the number of denoising iterations can be comprehensively set based on the local planning refresh cycle, edge controller computing power, and candidate trajectory coverage requirements. For example, when local planning needs to be completed within a short control cycle and real-time updates are emphasized, fewer initial noise sequences and fewer denoising iterations can be used. When the local scenario is more complex, with the influence of local boundary deformation and dynamic obstacle superposition, and when improved candidate trajectory coverage is required, the number of initial noise sequences and the number of denoising iterations can be appropriately increased. As an example, in a scenario with a local planning refresh cycle of 0.3s to 1.0s, 6 to 10 initial noise sequences can be selected, and 8 to 20 denoising iterations can be used to achieve a balance between real-time performance and candidate diversity. The above example values ​​are settings corresponding to the local planning refresh cycle and edge computing power conditions, and are not isolated parameter tuning results.

[0086] In one embodiment, the candidate obstacle avoidance trajectory can be represented as a sequence of discrete trajectory points within the future planning time domain, or as a sequence of control commands, a sequence of relative displacements, or a sequence of attitude changes. If represented as a sequence of discrete trajectory points, each candidate trajectory can contain multiple trajectory points arranged in chronological order, and each trajectory point includes at least planar position coordinates, and may also include orientation information if necessary. If represented as a sequence of control quantities, each candidate trajectory can consist of multiple velocity commands, steering commands, or attitude adjustment quantities at discrete moments. Any representation format can be used as long as it supports subsequent constraint checks, boundary violation judgments, and target trajectory selection.

[0087] In one embodiment, to ensure that the operational status information, task target information, obstacle occupancy information, and effective passage boundaries can be input into the diffusion strategy model in a form suitable for the narrow passage obstacle avoidance task of the button mushroom, a dual-branch conditional coding network can be used to form guiding condition features. The first branch processes the spatiotemporal environmental constraints within the local planning area, while the second branch processes the operational status information and task target information of the mobile robot. The outputs of the two branches are then fused to obtain a unified guiding condition feature.

[0088] In one implementation, the input to the first branch can be a spatiotemporal constraint tensor. Specifically, obstacle occupancy grids at each moment within the planning time domain can be stacked sequentially to form a multi-frame local occupancy grid sequence, and the effective passage boundary can be converted into a boundary constraint band corresponding one-to-one with the local occupancy grid at each moment. For the narrow passage scenario of button mushrooms, the local planning area usually exhibits a corridor-like geometric feature of "longer vertically and narrower horizontally". Therefore, the local occupancy grid preferably adopts a grid organization method with a resolution slightly higher along the extension direction of the passage than the horizontal resolution, so as to enhance the representation ability of the local boundary encroachment section and the dynamic obstacle passage direction. For example, in a scenario where the length of the local planning area is 2.4m to 4.8m and the width is 1.0m to 1.8m, the grid cell side length can be set to 20mm to 40mm. When the local boundary deformation of the button mushroom cultivation rack is minor and the robot needs to pass through with a small safety margin, 20mm can be used; when more attention is paid to the overall passage trend and the control refresh cycle is short, 30mm or 40mm can be used. If the planning time domain is set to 1.2s to 2.4s and the time offset step size is set to 0.1s, then 12 to 24 frames of spatiotemporal constraint tensors can be formed. The width of the boundary constraint band can be set according to the minimum safe passage margin, for example, it can be set to 5mm to 30mm; under the condition of transportation, in order to ensure conservative avoidance of local deformation boundaries, it can be taken as 20mm or 25mm.

[0089] In one embodiment, the first branch can employ a spatiotemporal convolutional coding subnetwork to extract features from the spatiotemporal constraint tensor. For example, a 3- to 5-layer convolutional coding structure can be used, where each convolutional kernel can be set to 3×3 or 3×5 to accommodate more pronounced occupancy changes along the channel direction in the narrow channel of the mushroom; the number of channels can be set to 16, 32, 64 or 32, 64, 128 respectively. For the scenario of this application, if the local planning area is small, the number of dynamic obstacles is small, and the real-time deployment at the edge is emphasized, a 3-layer convolutional coding structure can be used; if there are significant boundary deformations and complex disturbances such as mixed traffic of people and vehicles in the local area, a 4- or 5-layer convolutional coding structure can be used to improve the extraction capability of local boundary intrusion sections and dynamic occupancy changes. If necessary, a downsampling layer with a stride of 2 can also be set between adjacent convolutional layers to form multi-scale environmental constraint features.

[0090] In one embodiment, the input to the second branch can be a state target vector. The state target vector may include at least one of the following: the current position, current orientation, current linear velocity, current angular velocity, target position, and target orientation of the mobile robot. If the current task requires the robot to approach the picking area entrance or buffer channel connection area in a specific direction, a target entry direction or target parking posture representation can be further added. Considering that the dimension of the state target vector is usually low in the local transport scenario of button mushrooms, but the coupling between the state and the target has a significant impact on the trajectory positioning quality, in one embodiment, a multilayer perceptron coding subnetwork can be used to encode the state target vector. For example, 2 to 4 fully connected layers can be used, with hidden layer dimensions set to 64, 128, or 256. When the state target vector only includes basic information such as position, velocity, and target posture, a 2-layer fully connected layer with 128 dimensions per layer can be used; when further extended information such as transport status, area type, or target entry direction needs to be introduced, a 3-layer fully connected layer can be used, with hidden layer dimensions set to 128, 128, or 256. The above settings are not isolated parameter adjustments, but rather correspond to different conditional coding requirements for basic local obstacle avoidance scenarios and scenarios with operation alignment requirements.

[0091] In one implementation, the features from the two branches can be combined to form a unified guiding condition feature through concatenation, attention fusion, or conditional modulation. For example, the environmental constraint features output by the spatiotemporal convolutional coding sub-network can be pooled first, and then concatenated with the state target vector encoding result in the channel dimension to form a unified condition vector. In another implementation, a cross-attention mechanism can be used, with the state target feature as the query vector and the environmental constraint feature as the key-value pair, allowing the model to focus more on changes in obstacle occupancy and boundary constraints related to the current target's positioning direction and current operating state within the local planning region. For the narrow-channel scenario of the button mushroom, if edge deployment and inference speed are emphasized, channel-dimensional concatenation can be prioritized; if joint perception of target direction and boundary constraints in complex intersection areas is emphasized, cross-attention fusion can be used. By using the above-mentioned dual-branch conditional encoding method, on the one hand, it can maintain the local structural information of obstacle occupancy changes and effective passage boundaries in the time and space dimensions, and on the other hand, it can introduce the current operating state of the mobile robot and the target positioning requirements into the same conditional constraint space. This makes it more suitable for candidate trajectory generation tasks in the narrow passage scenario of mushroom cultivation, where there are dynamic obstacles, boundary intrusion and target alignment requirements at the same time.

[0092] In one embodiment, the denoising backbone network in the diffusion strategy model can adopt a one-dimensional temporal U-shaped network oriented towards trajectory sequences. Specifically, each initial noise sequence can be represented as a noise trajectory sample composed of N discrete trajectory points within the planning time domain, where each trajectory point may include planar position coordinates, and orientation information may be included if necessary. For example, when the planning time domain is 2.0s and the control update time resolution is 0.1s, N=20 can be taken; if the planning time domain is 1.5s and the time resolution is 0.1s, then N=15 can be taken. The denoising backbone network may include 3 to 4 levels of downsampling paths and corresponding upsampling paths, and the number of feature channels at each level can be set to 64, 128, 256 and 256, 128, 64 respectively. Guiding condition features can be injected into the feature layers of the denoising backbone network at each level through adaptive normalization, conditional bias injection, or cross-layer splicing, so that each step of denoising update not only depends on the current trajectory sample itself, but is also continuously constrained by obstacle occupancy information, effective passage boundaries, running status information, and task target information. For the scenario described in this application, a one-dimensional temporal U-shaped network is preferred for processing future short-term trajectory point sequences, because such networks can preserve the continuity of the trajectory as it unfolds over time, and also facilitate the introduction of conditional constraints in each layer while taking into account the efficiency of edge reasoning.

[0093] In one embodiment, the multi-step inverse diffusion denoising iteration can be performed as follows: First, construct corresponding initial trajectory samples for each initial noise sequence. Then, in each denoising iteration step, input the current trajectory sample and guiding condition features into the denoising backbone network, output the noise estimation result or trajectory correction amount corresponding to the current iteration step, and then update the current trajectory sample accordingly to obtain the trajectory sample for the next iteration step. For the narrow passage local obstacle avoidance task of Agaricus bisporus, a fast constraint check can be performed on the trajectory sample after each update step to check whether it has entered the obstacle-occupied area, whether it has exceeded the effective passage boundary, and whether it has violated the current execution constraints, and apply corrections to samples that deviate significantly from the feasible region. In this way, constraint information is not only used for final screening, but also participates in the formation process of candidate trajectories, thereby increasing the proportion of effective candidate trajectories generated in the end.

[0094] In one embodiment, the number of initial noise sequences and the number of denoising iteration steps should be set in conjunction with the control refresh cycle, edge computing power, and candidate coverage requirements in the scenario of this application, rather than being set in isolation. For example, when the local planning refresh cycle in a mushroom workshop is 0.3s to 0.8s and the vehicle controller needs to complete the generation and screening of candidate trajectories within one refresh cycle, 6 to 10 initial noise sequences can be selected, and 8 to 20 steps of reverse diffusion denoising iteration can be used to achieve a balance between real-time performance and candidate diversity. When the local scene is more complex, with obvious boundary deformation and dynamic obstacle superposition interference, and the vehicle platform has higher inference capabilities, the number of initial noise sequences can be appropriately increased to 10 to 16, and 15 to 25 steps of denoising iteration can be used. The above setting logic corresponds to the different requirements of "high-frequency local update scenario" and "improving candidate coverage capability in complex local scenario", rather than simply the result of model parameter tuning.

[0095] In one embodiment, the diffusion strategy model can be obtained through offline training. Training samples can come from historical passage trajectories, manually taught trajectories, and simulated trajectories generated in a digital twin environment within a button mushroom factory cultivation environment. Historical passage trajectories reflect boundary deformation, personnel movement, and local operational constraints in a real workshop; manually taught trajectories can supplement high-precision alignment or complex detour scenarios; and simulated trajectories can be used to expand sample coverage under different boundary intrusion degrees, different combinations of dynamic obstacles, and different area type switching conditions. To ensure that the training samples reflect the scenario characteristics of this application, in one embodiment, the training data preferably includes at least two of the following: edge passage scenarios of the button mushroom growth area, entrance passage scenarios of the picking area, and buffer channel connection scenarios. For example, the proportion of samples related to local boundary deformation in the training samples can be no less than 30%, the proportion of samples related to dynamic obstacle interaction can be no less than 30%, and the remaining samples can consist of ordinary narrow channel passage and target arrival scenarios. The basis for setting the above proportion is that the focus of this invention is on local obstacle avoidance generation under boundary deformation conditions. If there are too few relevant samples, the model will not be able to learn enough about the changes in real-world passage boundaries. The example proportion here is not a fixed requirement, but is used to illustrate that the training samples should cover the two key scenario factors of boundary deformation and dynamic obstacles.

[0096] In one embodiment, the offline training platform can be implemented using a workshop edge server or a research and development training workstation. For example, the training platform may include a multi-core CPU, at least 32GB of system memory, and a GPU with at least 8GB of video memory, used to jointly train the conditional coding network and the denoising backbone network. After training, the model parameters are deployed to the onboard industrial control computer of the mobile robot or the workshop edge control node for online inference. For the scenario described in this application, the onboard inference platform is preferably an industrial control platform with an embedded GPU or NPU, such as an onboard industrial control computer with an 8-core CPU, 16GB of memory, and 8GB of video memory, to ensure that candidate trajectory generation, screening, and subsequent control calls can be completed within the local planning refresh cycle in the narrow channel scenario of the button mushroom. By adopting a deployment method that combines offline training on an edge server with online inference on an onboard platform, both the sample expansion capability during the training phase and the real-time response capability during the runtime phase can be considered.

[0097] In one embodiment, supervised trajectory reconstruction loss and constraint consistency loss can be jointly used to optimize model parameters during training. The supervised trajectory reconstruction loss constrains the difference between the model-generated trajectory and the taught trajectory or historical passage trajectory, while the constraint consistency loss constrains the consistency between the generated trajectory and obstacle occupancy information, effective passage boundaries, and execution constraints. For this application, the constraint consistency loss preferably jointly measures the degree of boundary crossing, the degree of occupancy conflict, and the target positioning deviation. For example, when prioritizing improving the reliability of passage in boundary deformation scenarios, the weight of the boundary crossing consistency loss can be appropriately increased; when prioritizing improving the target alignment quality, the weight of the positioning deviation loss can also be appropriately increased. The above loss weight settings should match the emphasis on "passage safety" and "positioning stability" in the mushroom narrow passage local obstacle avoidance task.

[0098] In one implementation, the training samples for the diffusion strategy model can be constructed as "conditional feature-target trajectory" sample pairs. Specifically, multiple training samples can be extracted from historical travel trajectories, artificially taught trajectories, and digital twin simulation trajectories in the industrialized cultivation environment of button mushrooms. Each training sample includes: a guidance condition feature corresponding to a certain local planning scenario, and a target trajectory sequence corresponding to that guidance condition feature. The guidance condition feature can be formed by the combination of operating status information, operation target information, obstacle occupancy information, and effective passage boundaries; the target trajectory sequence can be represented by a discrete trajectory point sequence within the planning time domain, and each trajectory point includes at least planar position coordinates, and may also include orientation information if necessary.

[0099] In one embodiment, forward noise addition processing can be performed on the target trajectory sequence to generate noisy trajectory samples for the training phase. Specifically, Gaussian noise can be gradually superimposed onto the target trajectory sequence according to a noise scheduling sequence within a preset diffusion step count to obtain noisy trajectory samples with different noise intensities. The noise scheduling sequence can be set in a linear increment or in a scheduling method with slower initial growth and faster later growth. For example, the diffusion step count can be set to 50 to 200 steps; when the local planning refresh cycle is short and efficiency needs to be considered during the training phase, 100 steps can be used; when it is necessary to improve the model's ability to recover trajectory samples with different noise intensities, 150 or 200 steps can also be used. The above settings correspond to two types of training requirements: prioritizing training efficiency and prioritizing noise coverage.

[0100] In one embodiment, the input to the denoising backbone network during the training phase may include noisy trajectory samples, corresponding diffusion step identifiers, and guiding condition features, and the output may be the noise estimation result in the current noisy trajectory sample. That is, in this embodiment, a noise prediction training method is preferably adopted, enabling the network to learn to recover the noise components in the trajectory sequence under given condition features and current noise intensity. Subsequently, the noise prediction loss can be calculated based on the difference between the noise estimation result and the actual added noise; simultaneously, the trajectory reconstruction loss can be calculated based on the difference between the denoised and recovered trajectory and the target trajectory, and the constraint consistency loss can be calculated based on the degree to which the recovered trajectory satisfies obstacle occupancy information, effective passage boundaries, and execution constraints.

[0101] In one embodiment, the trajectory reconstruction loss can be calculated using the sum of squared positional deviations between the recovered trajectory and the target trajectory at each discrete time point, and may further include the endpoint attitude deviation; the constraint consistency loss may include at least one of the following: the amount of exceeding the effective passage boundary, the amount of intrusion into the obstacle-occupied area, and the amount of deviation from the target's position. For example, the three types of losses can be weighted and summed to form the total training loss. When prioritizing improving the reliability of passage under boundary deformation scenarios, the weight of the exceeding boundary loss can be increased; when prioritizing improving the target alignment quality, the weight of the position deviation loss can be increased. Each loss weight can be set according to the emphasis on "passage safety" and "position stability" in the mushroom narrow passage local obstacle avoidance task.

[0102] In one embodiment, the model parameters can be updated using a mini-batch iterative approach during training. For example, the batch size can be set to 16–64; 16 or 32 can be used when GPU memory resources are limited, and 48 or 64 can be used when training platform resources are sufficient. The number of training epochs can be set according to the convergence of the validation set loss, for example, it can be set to 50–200 epochs, and training stops when the validation set loss decreases below a preset threshold within several consecutive epochs. Through the above method, a post-training diffusion strategy model suitable for the narrow-channel local obstacle avoidance task of Agaricus bisporus can be obtained.

[0103] Regarding S105:

[0104] In one embodiment, multiple candidate obstacle avoidance trajectories can first undergo collision and boundary violation screening. For any candidate trajectory, it can be checked whether the trajectory enters an obstacle-occupied area within the planning time domain, whether it crosses the effective passage boundary in any local segment, and whether it exceeds the robot's current execution capabilities. Execution constraints refer to constraints related to the robot's actual operating capabilities, which may include maximum steering change rate, maximum lateral offset rate, maximum deceleration, minimum safe passage margin, or other limitations related to actual control execution. All of the above execution constraints can be obtained using conventional motion control constraint setting methods in the art.

[0105] In one embodiment, after eliminating obviously unexecutable candidate trajectories, a target obstacle avoidance trajectory can be determined from the remaining candidate trajectories. This determination process can be based on target positioning deviation, trajectory smoothness, travel distance, control cost, or a combination thereof. Those skilled in the art can select specific evaluation indicators according to actual task requirements. For narrow-channel mushroom transport scenarios, it is preferable to ensure that the target obstacle avoidance trajectory not only avoids boundary crossing and collision, but also maintains small attitude fluctuations and high tracking stability when approaching the target area. Through the above processing, this application can not only avoid generating local trajectories that do not match the actual boundaries based on the designed channel boundaries, but also obtain a target obstacle avoidance trajectory more suitable for actual execution when local boundary deformation occurs and is accompanied by dynamic obstacle interference.

[0106] Furthermore, in one embodiment, this application can be illustrated using a mushroom transport scenario. For example, a mobile robot needs to transport a vehicle along a buffer channel to a vicinity of the harvesting area. Within 3.5m in front of the robot, a cultivation rack boundary is detected to bulge inwards into the channel, while a worker is moving near the entrance of an adjacent channel. In this case, the area 4m in front of the robot, covering the current channel and the adjacent boundary buffer zone, can be defined as a local planning area. Dynamic obstacle occupancy information within this local planning area is extracted using environmental perception results, and the local deformation is obtained based on the difference between the observed boundary and the design boundary. If the boundary intrusion of this local section is 18mm, and the minimum safe passage margin requirement for the robot under the current transport condition is 25mm, then the effective passage boundary corresponding to this local section can be further corrected inwards into the channel. Then, the operating status information, work target information, obstacle occupancy information, and the corrected effective passage boundary are jointly encoded as guidance condition features. Based on multiple initial noise sequences, inverse diffusion denoising iteration is performed to generate multiple candidate trajectories. Candidate trajectories that exceed the corrected effective passage boundary, overlap with the area predicted by operators, or exhibit excessively drastic turning changes are eliminated. From the remaining trajectories, a trajectory close to the center of the passage, with minimal positioning deviation and high smoothness is selected as the target obstacle avoidance trajectory. This process maintains high passage safety and operational stability even when local boundary bulges and dynamic obstacles coexist.

[0107] It should be noted that the parameter ranges, threshold values, and numerical examples in the above embodiments are all used to illustrate the basis for the settings and applicable scenarios of this application, and do not constitute a limitation on the scope of protection of this application. Those skilled in the art can make equivalent adjustments to the above parameters based on factors such as robot size, channel width, carrying status, perception resolution, and control refresh cycle. As long as the core idea is still based on obstacle occupancy information within the local planning area and the effective passage boundary determined by the facility boundary position and local deformation, to guide the diffusion-type candidate trajectory generation process and complete the target trajectory screening, it should fall within the scope of protection of this application.

[0108] Optional, see Figure 2 The flowchart of a method for determining obstacle occupancy information provided in this application embodiment includes steps S201 to S202, wherein:

[0109] S201: Based on the current motion state of the dynamic obstacle, predict the predicted position of the dynamic obstacle at each moment in the planning time domain;

[0110] S202: Determine the obstacle occupancy information based on the predicted position at each time point.

[0111] In some scenarios, relying solely on the current position of dynamic obstacles for local obstacle avoidance can lead to the following situation: the candidate trajectory obtained by the robot during planning may be theoretically feasible, but by the time the robot actually reaches the corresponding position, the dynamic obstacle may have moved to the vicinity of that trajectory or even entered the corresponding area, causing the originally feasible candidate trajectory to become invalid during execution. This is especially true in the industrialized cultivation environment of button mushrooms, where workers, other mobile robots, and temporary transport vehicles may create dynamic disturbances near narrow passages, passage intersections, or buffer passage entrances. If obstacles are still treated as static targets, candidate trajectories are prone to frequent recalculation, affecting local traffic stability and efficiency. Therefore, in one embodiment, the predicted position of the dynamic obstacle at each moment within the planning time domain can be predicted based on its current motion state, and obstacle occupancy information can be determined based on the predicted position at each moment.

[0112] In one embodiment, the current motion state may include at least one of the following: the current position, current velocity, current direction of motion, and / or current acceleration of the dynamic obstacle. This information can be obtained using conventional dynamic target detection and tracking methods in the art, such as using lidar, depth cameras, millimeter-wave radar, or combinations thereof to perform target detection, target association, and state estimation on continuous environmental perception results, thereby obtaining the position change sequence of the dynamic obstacle in the local planning area and the corresponding velocity and direction information. For the purposes of this application, the specific category of the dynamic obstacle is not limited; any obstacle that may obstruct the robot's local passage within the planning time domain can be considered a dynamic obstacle and participate in subsequent prediction.

[0113] In one embodiment, the position of a dynamic obstacle at each moment within the planning time domain can be predicted short-term based on its current position and velocity information. For dynamic obstacles with relatively stable motion, a constant velocity extrapolation method can be used to obtain the predicted position at each moment; for dynamic obstacles with obvious acceleration / deceleration or turning trends, conventional short-time trajectory prediction methods in the art can also be used to obtain the predicted position at each moment. The predicted position at each moment refers to the predicted spatial position of the dynamic obstacle at each discrete moment within the planning time domain, which can be represented as two-dimensional planar coordinates, local grid position index, boundary envelope center point, or other positional expressions that can be used for subsequent occupancy determination. Therefore, the obstacle no longer participates in local obstacle avoidance planning based solely on its position at a single moment, but rather participates in local planning as a time-related position sequence.

[0114] In one implementation, the planning time domain can be set according to the robot's current speed, local channel length, and local obstacle avoidance response time requirements. For example, when performing precise passage or alignment operations in narrow channels, the planning time domain can be set to 1s to 3s; when it is necessary to determine yielding relationships in advance near local intersections or buffer channel entrances, the planning time domain can be appropriately extended to 2s to 4s. The time interval step length can be set according to the control refresh cycle and perception update frequency, for example, it can be set to 0.05s to 0.2s. If the controller refresh cycle is short and more detailed reflection of dynamic obstacle occupancy changes is required, 0.1s can be used; if more attention is paid to the overall traffic trend rather than high-frequency details, 0.2s can also be used. Through the above settings, the predicted position sequence of dynamic obstacles at multiple discrete moments can be obtained within the planning time domain.

[0115] In one embodiment, when determining obstacle occupancy information based on the predicted positions at each time point, the predicted positions of dynamic obstacles at each time point can be further expanded into corresponding occupancy areas. These occupancy areas can be determined based on the dynamic obstacle's dimensions, positioning error, prediction error, and safety margin. For example, for workers, the occupancy area can be determined based on the width of the human body's lateral contour and the normal walking swing range; for other mobile robots, the occupancy area can be determined based on their vehicle body dimensions, steering swing, and running error. The occupancy areas at each time point can be further combined to form a time-related occupancy sequence, serving as obstacle occupancy information within the local planning area. In this way, subsequent candidate trajectory generation and selection no longer rely solely on the obstacle distribution at the current time point, but can reflect the dynamic changes in obstacle occupancy within the planning time domain, thereby reducing the risk of temporal conflicts between candidate trajectories and dynamic obstacles during execution.

[0116] In other scenarios, even if the facility boundary locations and local deformation amounts within the local planning area are obtained, directly using the design boundary as the local passage boundary, or using a uniform contraction method to correct the entire passage boundary, can easily lead to two types of problems: First, overly conservative treatment of sections without significant deformation results in unnecessary compression of the passage space; second, insufficient correction of sections with local encroachment leads to the generated candidate trajectories still having a real risk of collision. To make the boundary conditions on which the local obstacle avoidance planning is based more closely resemble the actual passage conditions in the button mushroom cultivation environment, in one embodiment, the initial boundary of the passage can be determined based on the design boundary of the button mushroom cultivation rack, and the initial boundary of the passage can be corrected along the inner side of the passage according to the local deformation amount to obtain an effective passage boundary.

[0117] The design boundary refers to the edge position of the passageway corresponding to the mushroom cultivation rack in its design state, installation calibration state, or reference benchmark state. The design boundary can be pre-stored in a map model, equipment installation parameter file, area geometry template, or boundary calibration table, or it can be obtained through the initial mapping or installation calibration process. Determining the initial passageway boundary based on the design boundary means first determining the basic passageway boundary that the robot should refer to for local passage when there is no local deformation or when current local deformation correction is not considered. This initial passageway boundary can serve as a benchmark for subsequent boundary correction based on local deformation.

[0118] In one implementation, correcting the initial boundary of the channel along the inner side of the channel based on local deformation refers to selectively shrinking the boundary at locations within the local planning area that show a tendency to intrude into the channel. In other words, the initial boundary of the entire channel is not uniformly shrunken; instead, based on the actual observed local deformation, only the boundary segments affecting the robot's local passage are corrected. For segments with small local deformation that do not meet the boundary correction conditions, the original initial channel boundary can remain unchanged. For segments with large local deformation sufficient to compress the local passage margin, the boundary of that segment can be corrected towards the inner side of the channel according to the corresponding intrusion amount. The resulting effective passage boundary retains the passage capacity of normal sections while accurately reflecting passage restrictions in locally deformed sections.

[0119] In one implementation, when determining whether the local deformation meets the boundary correction condition, the threshold can be set by combining the robot's external dimensions, minimum safe passage margin, and current control accuracy. For example, the boundary correction threshold for local deformation can be set to 5mm to 40mm. When the passage is narrow and the robot is sensitive to the passage margin during transport, a smaller threshold, such as 8mm or 10mm, can be used; when the passage is relatively wide and the robot's operating speed is low, a larger threshold, such as 20mm, can be used. If the local deformation of a certain segment exceeds the boundary correction threshold, it can be determined that the segment needs passage boundary correction; otherwise, the initial passage boundary of the segment can be retained. The above example values ​​correspond to different boundary correction requirements in narrow-margin, high-risk passage scenarios and relatively wide-margin, low-speed passage scenarios, respectively.

[0120] In one implementation, after boundary correction, the corrected boundary segment can be continuously spliced ​​with the uncorrected boundary segment to form a complete and effective passage boundary. Through this approach, the boundary upon which local obstacle avoidance planning is based is no longer the ideal boundary under design conditions, but rather reflects the actual boundary conditions after local deformation of the mushroom cultivation rack. Thus, when generating and screening candidate trajectories within the local planning area, the execution risk arising from a theoretical passage width meeting the passage conditions but insufficient actual passable range in the local area can be effectively avoided.

[0121] In one embodiment, to avoid abrupt boundary changes at the junction of the corrected and uncorrected segments, the boundary correction amount can be smoothly transitioned using linear interpolation or cubic spline interpolation within a range of 2 to 3 boundary sampling points before and after the junction. This transition process ensures that the effective passage boundary after splicing remains geometrically continuous, which is more conducive to the stable execution of subsequent candidate trajectory generation and boundary constraint judgment.

[0122] In other scenarios, even if obstacle occupancy information and effective passage boundaries are determined, if they are not continuously applied to each denoising update step during candidate trajectory generation but are only used as post-selection conditions, a large number of generated trajectory samples may still be concentrated in areas that do not match the actual passage boundaries, resulting in a low proportion of usable candidate trajectories, increasing the burden of subsequent selection, and affecting real-time performance. Especially in the narrow passage scenario of the mushroom agaric mushroom, if candidate trajectories are not sufficiently constrained during the generation stage, trajectory samples that are too close to the edge, cross the boundary, or have excessive local turning are likely to be generated. To enable candidate trajectories to gradually converge to a trajectory distribution that meets the current constraints during the generation process, in one embodiment, the multi-step inverse diffusion denoising iteration includes: during the inverse diffusion denoising iteration process, guiding the update of trajectory samples based on obstacle occupancy information, effective passage boundaries, and preset execution constraints.

[0123] In this process, guided updates refer to updating the trajectory samples not only based on the denoising results of the current trajectory samples in each iteration of the inverse diffusion denoising, but also by incorporating obstacle occupancy information, effective passage boundaries, and execution constraints. This constrains the update direction of the trajectory samples, causing them to gradually converge towards an executable trajectory distribution that better meets the requirements of the scenario. In other words, the multi-step inverse diffusion denoising in this application does not unconditionally recover the trajectory from random noise, but rather introduces constraint information related to the current scenario in each step, thereby reducing the probability of the generated results not matching the local real passage conditions.

[0124] In one embodiment, obstacle occupancy information plays a role in guided updates by preventing trajectory samples from entering the occupied areas of dynamic or static obstacles within the planning time domain; effective passage boundaries play a role in guided updates by preventing trajectory samples from exceeding the actual passable range within the current local planning area; and execution constraints play a role in guided updates by preventing trajectory samples from exhibiting state changes that exceed the robot's current control capabilities.

[0125] Execution constraints refer to the constraints related to the actual operational capabilities of the mobile robot in the current scenario, and their scope is as described above. In this embodiment, execution constraints may specifically include at least one of the following: maximum rate of change of steering, maximum rate of lateral deviation, maximum deceleration, allowable attitude deviation when approaching the target area, and minimum safe passage margin relative to the effective passage boundary. For the narrow passage transport scenario of button mushrooms, execution constraints are preferably set in association with the actual passage boundary, rather than simply based on general kinematic limits.

[0126] In one embodiment, the minimum safe passage margin can be set comprehensively based on the robot's external dimensions, the lateral swing range during transport, and the degree of local boundary intrusion. For example, it can be set to 10mm to 40mm; when the local deformation of the passage boundary is significant, the robot is in transport mode, and safe passage must be prioritized, it can be set to 25mm; when the local passage is relatively straight and the robot passes through with precise alignment at low speed, it can also be set to 15mm. The above example values ​​correspond to different execution requirements, namely, prioritizing safety and prioritizing precise passage.

[0127] In one embodiment, the number of steps in the multi-step inverse diffusion denoising iteration can be set comprehensively based on the controller's real-time requirements, scene complexity, and trajectory smoothness requirements. When the local channel is short, there are few dynamic obstacles, and the system emphasizes rapid updates, fewer denoising steps can be used; when there are obvious boundary deformations and dynamic interferences in the local channel, but it is necessary to improve the quality of candidate trajectories, the number of denoising steps can be appropriately increased. By continuously introducing the above-mentioned constraint information in each denoising update step, the multiple candidate obstacle avoidance trajectories obtained can be more concentratedly distributed in areas that do not conflict with obstacle occupancy, do not exceed the effective passage boundary, and satisfy the execution constraints, thereby improving the effectiveness of subsequent trajectory selection and the overall local planning efficiency.

[0128] Furthermore, in one embodiment, the above process can be illustrated using a mushroom cultivation scenario. For example, a mobile robot operates along the connecting channel between the harvesting area and the buffer channel. A worker is detected moving at approximately a constant speed along the channel intersection in front of its local planning area, while the boundary of the cultivation rack on one side of the channel shifts inwards in a local section. At this point, the worker's speed and direction of movement can be determined based on the worker's current position over two or more consecutive frames. Then, the predicted positions at each discrete moment are extrapolated within the planning time domain to form corresponding obstacle occupancy information. On the other hand, the current observed boundary and the design boundary can be compared under a unified local planning coordinate system to obtain the local deformation. Only boundary sections that meet the boundary correction conditions are corrected along the inner side of the channel to form an effective passage boundary. Then, in the multi-step inverse diffusion denoising process, the obstacle occupancy information, the effective passage boundary, and the execution constraints are applied together to update the trajectory samples, enabling the obtained candidate trajectories to actively avoid the worker's future occupancy area while avoiding proximity to the local intrusion boundary. In this way, the effective proportion of candidate trajectories and the efficiency of subsequent screening can be improved when local boundary deformation and dynamic obstacles coexist.

[0129] It should be noted that the dynamic obstacle prediction method, boundary correction threshold, time step size, and denoising iteration steps in the above embodiments can all be conventionally adjusted by those skilled in the art based on robot size, site scale, control refresh cycle, and real-time requirements. As long as it still reflects the determination of obstacle occupancy information based on the predicted position of dynamic obstacles at each moment in the planning time domain, the correction of the initial boundary of the channel based on local deformation to form an effective passage boundary, and the guidance and updating of trajectory samples based on obstacle occupancy information, effective passage boundary, and execution constraints during the reverse diffusion denoising process, it should be considered as the technical solution described in this application.

[0130] Optional, see Figure 3 The flowchart of a method for determining a target obstacle avoidance trajectory provided in this application embodiment includes steps S301 to S302, wherein:

[0131] S301: Sort the remaining candidate obstacle avoidance trajectories according to the target positioning deviation and smoothness;

[0132] S302: The candidate obstacle avoidance trajectory with the best ranking result is determined as the target obstacle avoidance trajectory.

[0133] In some scenarios, even after eliminating obviously unexecutable candidate obstacle avoidance trajectories through obstacle occupancy information, effective passage boundaries, and execution constraints, the remaining candidate trajectories may still simultaneously meet the basic passage conditions. However, these candidate trajectories differ in terms of positioning accuracy, turning smoothness, and subsequent trackability. If only one remaining trajectory is randomly selected, or if a trajectory is selected based solely on a single length index, the following situation may easily occur: although the trajectory can avoid obstacles and remain within the effective passage boundary, it may exhibit significant attitude deviations, overly drastic turning changes, or heavy tracking control burdens when approaching the target position, thereby affecting the positioning stability and overall passage efficiency in the narrow passage scenario of the mushroom tree. Based on this, in one embodiment, the remaining candidate obstacle avoidance trajectories can be sorted according to the target positioning deviation and smoothness, and the candidate obstacle avoidance trajectory with the best sorting result can be determined as the target obstacle avoidance trajectory.

[0134] The target positioning deviation refers to the degree of deviation between the endpoint state of a candidate obstacle avoidance trajectory and the target operational requirements. It can include at least endpoint position deviation, endpoint orientation deviation, or a combination of both. For scenarios requiring entry into harvesting areas, buffer channel docking positions, or specific transfer stations, simply reaching the vicinity of the target area is usually insufficient; the robot also needs to enter the target position in a suitable posture to facilitate subsequent alignment, handover, or continued passage. Therefore, the distance between the endpoint position and the target position of each remaining candidate obstacle avoidance trajectory can be calculated, and the deviation between the trajectory endpoint orientation and the target entry direction or target parking posture can be further calculated to form the corresponding target positioning deviation.

[0135] In one embodiment, the endpoint position deviation can be represented by Euclidean distance or the deviation after decomposition along the target approach direction, and the endpoint orientation deviation can be represented by the angle difference between the endpoint heading angle and the target heading angle. Both the endpoint position deviation and the endpoint orientation deviation can be obtained using conventional geometric calculation methods in the art. To facilitate the joint sorting of deviations of different dimensions, the endpoint position deviation and the endpoint orientation deviation can be further normalized to form a comprehensive positioning deviation index.

[0136] For example, when performing a transfer and alignment task within a narrow channel of a button mushroom, if more emphasis is placed on precise positioning, the weight of the endpoint position deviation in the overall index can be appropriately increased; if more emphasis is placed on the consistency of attitude upon entering the target area, the weight of the endpoint orientation deviation can also be increased. This application does not limit the specific weighting form, as long as it can reflect the degree of matching between the candidate trajectory endpoint state and the target operation requirements.

[0137] In one embodiment, the smoothness referred to in this application is used to characterize the smoothness of the candidate obstacle avoidance trajectory in terms of spatial morphology or control changes. For local obstacle avoidance trajectories, if the trajectory has abrupt turns, excessive curvature changes, or discontinuous directional changes in local sections, although it may meet obstacle avoidance and boundary constraints, it is prone to problems such as increased lateral sway of the vehicle body, excessively rapid changes in control commands, or unstable transport status during actual execution. Especially in the mushroom transport scenario, if the robot makes overly drastic steering adjustments near the local deformation boundary in a narrow channel, it may reduce the stability of passage. Therefore, the smoothness of the trajectory can be characterized by at least one of the following: the rate of change of direction, the rate of change of curvature, the change of the angle between adjacent trajectory segments, or the amplitude of the change in control quantity on the discrete trajectory point sequence of the candidate obstacle avoidance trajectory.

[0138] In one implementation, each candidate obstacle avoidance trajectory can be discretely sampled to obtain multiple trajectory points arranged in chronological order. Then, based on the changes in the direction of the lines connecting adjacent trajectory points, the local turning change is calculated, and a smoothness evaluation index is formed accordingly. For example, in one implementation, the sum of the absolute values ​​of the tangential angles between adjacent trajectory segments can be used as the smoothness cost; in another implementation, the cumulative value of the curvature changes corresponding to adjacent discrete points can also be used as the smoothness cost. The larger the index value, the more pronounced the local trajectory turning, the more frequent the control switching, and the worse the smoothness. For scenarios with short local channel lengths but high alignment accuracy requirements, a moderate increase in local turning changes can be allowed to achieve smaller positioning deviations; for scenarios with large channel boundary undulations but requiring stable passage, the priority of smoothness can be appropriately increased.

[0139] In one embodiment, when sorting the remaining candidate obstacle avoidance trajectories, corresponding evaluation values ​​can be obtained based on the target positioning deviation and smoothness, and then a comprehensive ranking index can be formed according to a preset method. For example, weighted summation, hierarchical sorting, or priority sorting methods can be used for comprehensive evaluation. For instance, in one embodiment, the trajectories can be initially sorted by target positioning deviation from smallest to largest, and then re-sorted by smoothness from best to worst among candidate trajectories with similar deviations; in another embodiment, a comprehensive score can be calculated first, and then the trajectories can be directly sorted based on the comprehensive score. This avoids simply pursuing the shortest trajectory or the minimum local detour while ignoring the target positioning quality and execution smoothness.

[0140] In one implementation, parameters related to target positioning deviation and smoothness can be set according to the task type, channel geometry, and control precision. For example, the allowable range for endpoint position deviation can be set based on the alignment precision requirements of the target workstation, such as 10mm to 80mm; 20mm can be used when precise entry into the picking area or buffer channel connection area is required; and 50mm can be used when stable passage through a local area is required without precise alignment. The allowable range for endpoint orientation deviation can be 1° to 12°; 3° to 5° can be used for scenarios requiring entry into the target area in a specified direction; and can be appropriately relaxed for general passage scenarios. The above example values ​​correspond to different setting requirements for precise positioning scenarios and general local passage scenarios, respectively. Through the above sorting process, the finally selected target obstacle avoidance trajectory not only has obstacle avoidance feasibility but is also more conducive to actual execution and target positioning.

[0141] In some scenarios, if local obstacle avoidance planning is always performed on a fixed area, two types of problems can easily arise: First, if the local planning area is too large, it will include areas that do not require local detour processing, increasing the computational burden; second, if the local planning area is too small, it may fail to cover the critical areas that truly affect current passage or arrival operations, causing the planning results to quickly become invalid after leaving the local window. Especially in the industrialized cultivation environment of button mushrooms, robots usually have predetermined travel trajectories, such as running along the center line of existing channels, predetermined transport paths, or travel routes issued by the scheduling system. In most normal sections, the passage requirements can be met simply by following the predetermined travel trajectory; however, during the stages of local boundary deformation, dynamic obstacle intervention, or approach to the target workstation, local obstacle avoidance planning needs to be re-performed. Based on this, in one embodiment, areas where passage or arrival operations cannot be completed solely by following the predetermined travel trajectory under the current environmental conditions can be identified as local planning areas.

[0142] In one embodiment, the predetermined travel trajectory referred to in this application refers to the basic travel path determined by the robot based on a map, scheduling results, workstation tasks, or preset transportation rules before triggering local obstacle avoidance reconfiguration. This predetermined travel trajectory can be a channel center reference line or a continuous path segment provided by the upper-level planning module.

[0143] This application does not limit the specific source of the predetermined travel trajectory, as long as it can serve as the basic path for the robot to perform passage or positioning operations under normal circumstances. In one embodiment, it can be first determined whether the robot will conflict with obstacle occupancy information, exceed the effective passage boundary, or fail to meet the current target positioning requirements when it continues to run along the predetermined travel trajectory. If any of the above situations exist, it can be determined that the corresponding area cannot complete the passage or positioning operation based solely on the predetermined travel trajectory, and it can be identified as a local planning area.

[0144] In one embodiment, the local planning area can be determined as follows: starting from the robot's current position, a path segment within a predetermined length range is searched forward along a predetermined travel trajectory. It is then checked whether this path segment overlaps with the currently occupied obstacle area, enters a non-passable area obtained after boundary correction, or whether its endpoint posture fails to meet the current operational objective requirements. If any of the above situations are detected, the area containing the abnormal path segment and its preceding and following buffer ranges can be determined as the local planning area. The preceding and following buffer ranges can be set according to the robot's braking distance, posture adjustment distance, and control response time. For example, the forward buffer length can be set to 0.5m to 2m, and the backward buffer length can be set to 0.2m to 1m. When the forward posture needs to be adjusted in advance and detour space is reserved, a larger forward buffer length can be used; when only rapid processing of nearby local obstacles is required, a smaller value can be used.

[0145] In one implementation, in addition to traffic conflicts, the requirements for reaching the target can also serve as an important basis for determining the local planning area. For example, in the mushroom picking area or the buffer channel connection area, although a predetermined travel trajectory can guide the robot to the vicinity of the target, continuing to run solely according to the predetermined trajectory may not meet the requirements of the specified entry direction or destination posture. In this case, the area near the target that requires posture adjustment and precise positioning can also be determined as the local planning area. Thus, the local planning area is not limited to the detour area in the sense of obstacle avoidance, but can also be a precise operation area where the target positioning task cannot be completed solely by relying on the predetermined travel trajectory.

[0146] In one implementation, the size of the local planning region can be dynamically determined based on the current task type and environmental complexity. For scenarios with minor local boundary deformation and few obstacles, a smaller local planning region can be used to improve planning response speed; for scenarios with significant boundary encroachment, dense dynamic obstacles, or high target alignment requirements, the local planning region can be appropriately expanded to form more stable candidate trajectories in advance. Through this approach, the local planning region is no longer a fixed window, but a local planning range adapted to the current environmental state and task requirements, thus balancing planning efficiency and effectiveness.

[0147] In some scenarios, even if a target obstacle avoidance trajectory has been selected, if the complete trajectory is generated only once and executed directly in an open loop, the originally selected trajectory may gradually become invalid as dynamic obstacles continue to move, boundary states change further, or deviations occur during robot execution. Especially in the narrow passage environment of mushroom cultivation, where local clearance is limited, if the robot continuously accumulates tracking deviations during execution, or if the occupancy relationships of surrounding dynamic obstacles change, the original trajectory may no longer meet the current clearance conditions. Recalculating the entire trajectory after a significant conflict not only increases the discontinuity of trajectory execution but may also reduce local clearance efficiency. Therefore, in one embodiment, the mobile robot can be controlled to execute the target obstacle avoidance trajectory in a rolling execution mode, and during execution, in response to the current trajectory no longer meeting clearance conditions or the deviation of the current position from the current trajectory meeting replanning conditions, the target obstacle avoidance trajectory is regenerated based on the updated current state.

[0148] In one embodiment, the rolling execution method referred to in this application means that instead of issuing and continuously executing the entire target obstacle avoidance trajectory at once, the target obstacle avoidance trajectory is divided into continuously updated execution windows. Within each execution window, only the trajectory segment or control command required for the current time period is issued, and the applicability of subsequent trajectories is reassessed based on the latest state after the execution window progresses. This allows the robot to continuously maintain its responsiveness to changes in the local environment while executing the target trajectory. The length of the execution window can be set according to the control cycle, robot speed, and rate of environmental change. For example, in one embodiment, the execution window can be set to 0.3s to 1.5s; 0.5s can be used in scenarios with frequent dynamic obstacles and high-frequency updates; and about 1.0s can be used in scenarios with relatively stable channels and low-speed robot alignment.

[0149] In one embodiment, the phrase "the current execution trajectory no longer meets the passage conditions" as used in this application means that the subsequent trajectory segment corresponding to the originally selected target obstacle avoidance trajectory at the current moment no longer meets the local passage requirements. For example, the predicted position of a dynamic obstacle changes, causing the original trajectory to overlap with the new obstacle-occupied area; another example is that after the local boundary state is updated, some trajectory segments of the original trajectory have approached or exceeded the new effective passage boundary; yet another example is that when the robot approaches the target area, the original trajectory can no longer meet the posture or entry direction required for the current positioning operation. When any of the above situations occur, it can be considered that the current execution trajectory no longer meets the passage conditions.

[0150] In one embodiment, the deviation of the current position from the current execution trajectory, as referred to in this application, satisfies the replanning condition when the deviation between the robot's actual operating state and the target obstacle avoidance trajectory reaches a preset replanning threshold. For example, whether the replanning condition is triggered can be determined based on the lateral deviation, longitudinal deviation, orientation deviation, or a combination thereof from the robot's current position to the reference point of the current execution trajectory. Exemplarily, the lateral deviation threshold can be set to 10mm–60mm; in scenarios with narrow passages and small boundary margins, it can be 15mm or 20mm; in relatively wide passages, it can be appropriately widened. The orientation deviation threshold can be set to 1°–8°, specifically determined based on the current operating speed and passage width. These thresholds correspond to different deviation tolerance requirements for high-precision narrow passage scenarios and general local passage scenarios.

[0151] In one implementation, when a replanning condition is triggered, the target obstacle avoidance trajectory can be regenerated based on the updated current state. The updated current state can include at least one of the following: the robot's new operating state information, recalculated obstacle occupancy information, updated local planning area, and newly determined effective passage boundaries. In other words, replanning is not simply fine-tuning the trajectory under the original scenario assumptions, but rather re-executing the local obstacle avoidance planning process with the latest state to obtain a new target obstacle avoidance trajectory that better reflects the current real-world environmental conditions. This avoids delays in replanning results caused by using old state information.

[0152] Furthermore, in one embodiment, the above process can be illustrated using a narrow passage transport scenario for button mushrooms. For example, a mobile robot traverses a buffer passage along a target obstacle avoidance trajectory, executing the control segment corresponding to the first 0.6 seconds of the target trajectory within the first execution window. During execution, it is detected that the speed of the worker ahead has decreased compared to the original prediction, causing the yield window corresponding to the subsequent trajectory to shorten. Simultaneously, the robot itself experiences a certain lateral offset due to local ground disturbances. At this point, it can be determined that the subsequent part of the original target obstacle avoidance trajectory no longer meets the current passage conditions, and the lateral deviation of the robot's current position relative to the current execution trajectory is close to a preset threshold. Therefore, a new target obstacle avoidance trajectory is regenerated based on the updated current position, obstacle occupancy information, and effective passage boundary, and execution is switched in the next execution window. By combining this rolling execution with condition-triggered replanning, the continuity, stability, and environmental adaptability of the local obstacle avoidance process can be maintained even when local boundary deformation and dynamic obstacle changes coexist.

[0153] It should be noted that the rolling execution window, deviation threshold, and replanning triggering conditions in the above embodiments can all be conventionally set by those skilled in the art in combination with the control refresh cycle, robot outline dimensions, channel width, and scene dynamics. As long as it still reflects the determination of areas that cannot complete passage or positioning operations solely based on a predetermined travel trajectory as local planning areas, and the maintenance of the continuous effectiveness of local obstacle avoidance trajectories through rolling execution and condition-triggered replanning during the execution of the target trajectory, it should be considered a technical solution described in this application.

[0154] In some scenarios, even if several executable candidate trajectories have been identified based on obstacle occupancy information, effective passage boundaries, and execution constraints, these candidate trajectories may be similar in terms of geometrical feasibility and basic execution capabilities, but their impact on the growth environment of button mushrooms will differ. Especially in the factory cultivation environment of button mushrooms, the temperature and humidity conditions often differ between the growth zone, harvesting zone, and buffer passage. If a mobile robot crosses the boundary of an area with significant temperature and humidity differences at a high speed along a candidate passage direction, it may cause local airflow disturbances, intensified heat and humidity exchange, or an overly abrupt environmental transition, thereby affecting the stability of the button mushroom growth environment. Ranking candidate trajectories solely based on positioning deviation and smoothness makes it difficult to distinguish between geometrically feasible candidate trajectories with different degrees of environmental disturbance. Based on this, in one embodiment, the regional type of the local planning area in the mushroom growth area, harvesting area or buffer channel can be further determined, and the regional temperature and humidity transition parameters of the local planning area along the candidate passage direction can be determined; according to the cross-regional path, running speed and regional temperature and humidity transition parameters corresponding to the candidate obstacle avoidance trajectory, the regional environmental disturbance cost is determined, and the regional environmental disturbance cost is used as one of the sorting criteria when determining the target obstacle avoidance trajectory.

[0155] In one embodiment, the "region type" referred to in this application refers to the functional zoning attribute of a local planning area in the button mushroom production process. The region type can be pre-stored in a workshop digital map, region identification table, or scheduling system region configuration file, or it can be obtained by mapping the robot's current position and candidate trajectory through the region to the workshop zoning model. For example, the growth area can correspond to an area prioritizing stable control of the button mushroom growth environment, the harvesting area can correspond to an area with relatively frequent personnel operations and dynamic changes in traffic flow, and the buffer channel can correspond to an area that undertakes the functions of connecting and transferring between regions. Different region types not only represent different locations but also reflect the sensitivity of the area to temperature and humidity disturbances and the differences in preferred traffic strategies. For example, the growth area is usually more sensitive to environmental fluctuations, the harvesting area focuses more on the coordinated passage of people and vehicles, and the buffer channel can undertake a greater environmental transition buffer function.

[0156] In one embodiment, the "regional temperature and humidity transition parameter" referred to in this application is a parameter used to characterize the degree of temperature and humidity environmental change and the smoothness of the transition along the candidate travel direction in a local planning area. This parameter is not simply a temperature or humidity value at a single moment, but rather a comprehensive measure reflecting the magnitude of environmental differences and the gradient of these differences along the candidate travel direction. Specifically, several environmental sampling points can be extracted along the candidate obstacle avoidance trajectory or its corresponding travel direction, and the temperature and humidity values ​​corresponding to each sampling point can be obtained. The regional temperature and humidity transition parameter is then determined based on the changes in these temperature and humidity sampling values ​​along the candidate travel direction.

[0157] In one embodiment, environmental sampling points can be determined based on the length of the local planning area and the density of temperature and humidity sensors. If area-level environmental sensors are already installed in the workshop, data from adjacent area sensors can be directly retrieved and interpolated according to the candidate travel direction. If the robot itself is equipped with a temperature and humidity detection module, it can also sample the local environment during operation and form a temperature and humidity sequence along the candidate travel direction. To ensure that the regional temperature and humidity transition parameters accurately reflect the environmental change trend along the candidate travel direction, the sampling interval can be set in conjunction with the channel length, sensor resolution, and the smoothness of environmental changes. For example, the sampling interval can be set to 0.1m to 1.0m; when the area boundary is short and the environmental differences change rapidly, 0.2m or 0.3m can be used; when the buffer channel is long and the environmental transition is relatively smooth, 0.5m or 0.8m can be used. The above example values ​​correspond to different sampling requirements for short-distance rapid transition scenarios and long-distance smooth transition scenarios, respectively.

[0158] In one embodiment, the temperature and humidity differences between adjacent sampling points along the candidate travel direction, along with the corresponding transition distances, can be combined to calculate the regional temperature and humidity transition parameters. For example, in one embodiment, the temperature and humidity change rates between adjacent sampling points can be calculated separately, and then multiple change rates can be accumulated or weighted to form a comprehensive parameter characterizing the severity of environmental transition along the candidate travel direction. In another embodiment, the regional temperature and humidity transition parameters can also be formed directly based on the temperature and humidity difference between the starting and target areas, combined with the transition distance between them. This application does not limit the specific mathematical expression of this parameter, as long as it reflects the relationship that the greater the environmental difference and the shorter the transition distance along the candidate travel direction, the more severe the corresponding environmental transition. That is, the regional temperature and humidity transition parameters are preferably positively correlated with the degree of temperature and humidity difference and negatively correlated with the transition distance or transition buffer length.

[0159] In one embodiment, to facilitate ranking and comparison among different candidate trajectories, the regional temperature and humidity transition parameters can be dimensionless or normalized. For example, they can be mapped to a transition intensity index between 0 and 1, or an environmental transition score between 0 and 100. If a dimensionless index between 0 and 1 is used, the index can approach 0 when the temperature and humidity difference along the candidate travel direction is small and the transition distance is long; and it can approach 1 when the temperature and humidity difference along the candidate travel direction is large and the transition distance is short. The purpose of this normalization method is to facilitate its inclusion in the calculation of regional environmental disturbance costs along with subsequent parameters such as operating speed and cross-regional path length, rather than limiting a specific score range.

[0160] In one embodiment, the "regional environmental disturbance cost" referred to in this application is an evaluation metric used to characterize the impact of candidate obstacle avoidance trajectories on the stability of the mushroom growth environment and the smoothness of cross-regional environmental transitions. This evaluation metric does not directly characterize geometric obstacle avoidance feasibility, but is used to distinguish the differences in environmental disturbance risk caused by different candidate trajectories when passing through different regional boundaries. Generally speaking, if the cross-regional path of the candidate obstacle avoidance trajectory is long, the robot's operating speed is high, and the corresponding regional temperature and humidity transition parameters are large, it means that the candidate trajectory may cross regional boundaries with significant environmental differences more quickly and over a larger area, thus corresponding to a higher regional environmental disturbance cost; conversely, if the cross-regional path of the candidate trajectory is short, the operating speed is low, and the environmental transition is relatively smooth, its regional environmental disturbance cost can be relatively low.

[0161] In one embodiment, the regional environmental disturbance cost can be determined based on the cross-regional path, operating speed, and regional temperature and humidity transition parameters corresponding to the candidate obstacle avoidance trajectory. The "cross-regional path" referred to in this application refers to a path segment in the candidate obstacle avoidance trajectory that crosses the boundaries of different regional types or crosses areas with significant environmental differences along the candidate travel direction. The operating speed can be represented by the average operating speed, peak operating speed, or local operating speed of the boundary passage section corresponding to the candidate trajectory. In one embodiment, the cross-regional path length, boundary passage speed, and regional temperature and humidity transition parameters can be combined in a preset manner to form the regional environmental disturbance cost. For example, weighted summation, product combination, or hierarchical scoring methods can be used for calculation. If more emphasis is placed on the speed factor, the weight corresponding to the operating speed can be increased; if more emphasis is placed on the regional environmental differences themselves, the weight corresponding to the regional temperature and humidity transition parameters can be increased. It should be noted that the relevant weights or scoring intervals can be set according to the workshop environmental control requirements, and their setting basis should correspond to the environmental stability requirements. For example, when environmental control requirements in the growing area are strict, the influence weight of regional temperature and humidity transition parameters can be increased; when personnel work frequently in the picking area and need to maintain high traffic efficiency, the proportion of environmental disturbance costs in the overall ranking can be appropriately reduced.

[0162] In one embodiment, an environmental sensitivity coefficient can be further set for the regional environmental disturbance cost to reflect the different tolerances of different regional types to environmental disturbances. For example, the environmental sensitivity of the growth zone can be set higher than that of the harvesting zone and the buffer channel, making the candidate trajectory more sensitive to the regional environmental disturbance cost when crossing the boundary of the growth zone. The setting of the environmental sensitivity coefficient can be determined according to the environmental stability requirements of different production areas of button mushrooms. For example, when it is necessary to prioritize minimizing local temperature and humidity fluctuations in the growth zone, the growth zone can correspond to a higher sensitivity coefficient; when the buffer channel itself is a regional transition space, its sensitivity coefficient can be appropriately reduced. The above setting logic makes the regional environmental disturbance cost no longer a simple mathematical score, but directly related to the actual environmental control objectives of the button mushroom factory cultivation scenario.

[0163] In one embodiment, when determining the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories, the regional environmental disturbance cost can be used as one of the ranking criteria, participating in the candidate trajectory ranking along with the target positioning deviation and smoothness. For example, candidate trajectories with environmental disturbance costs exceeding a preset upper limit can be eliminated first, and then the remaining trajectories can be comprehensively ranked based on the target positioning deviation, smoothness, and regional environmental disturbance cost; alternatively, the regional environmental disturbance cost can be directly combined with the target positioning deviation and smoothness to form a comprehensive score. For local planning tasks near the button mushroom growth area, it is preferable to give the regional environmental disturbance cost a higher weight in the comprehensive ranking to balance traffic safety and growth environment stability; for short-distance traffic tasks near the harvesting area, its proportion can be appropriately reduced to balance traffic efficiency and alignment accuracy.

[0164] Furthermore, in one embodiment, the above process can be illustrated with a specific scenario. For example, a mobile robot needs to move from a buffer channel to a transfer interface area adjacent to the growth area. The system determines the type of the local planning area, including the area near the boundary between the buffer channel and the growth area, based on the map markers of the local planning area, and extracts multiple temperature and humidity sampling points along the candidate travel directions. If a significant temperature and humidity difference is found in the candidate directions, and one of the candidate trajectories approaches the growth area at a higher speed along a longer cross-area path, the environmental disturbance cost corresponding to that candidate trajectory can be rated as high; another candidate trajectory, although slightly longer, has a shorter cross-area segment and a lower speed, thus having less impact on the environmental stability of the growth area, and its environmental disturbance cost can be rated as low. Finally, when ranked together with the target positioning deviation and smoothness, the latter is more likely to be selected as the target obstacle avoidance trajectory. In this way, local obstacle avoidance planning can not only focus on geometric travel and trajectory smoothness, but also take into account the stability of the mushroom growth environment.

[0165] In some scenarios, uniformly correcting the initial boundary of a passageway based solely on boundary offsets on a two-dimensional plane can easily overlook the following: Mushroom cultivation racks typically have multi-layered structures, and the actual spatial envelope height affecting passage judgment differs between unloaded and loaded conditions. Furthermore, local deformations of the cultivation rack boundary are not uniform across the entire height range, but may only intrude into the passageway in certain local sections corresponding to specific height layers. If boundary deformations related to the robot's current passage envelope height are not distinguished, boundary corrections may be applied to high or low boundary changes that do not actually affect robot passage, resulting in unnecessary compression of the passageway; or, locally intrusive boundaries that are truly within the robot's current passage envelope height may be missed, leading to a mismatch between the planning results and the actual space. Based on this, in one embodiment, the passage envelope height range corresponding to the mobile operation robot under the carrying condition can be determined first; within the passage envelope height range, the deformed intrusion section of the mushroom cultivation rack that intrudes into the inner side of the channel relative to the initial boundary of the channel and the intrusion amount corresponding to each deformed intrusion section can be determined; then, according to the intrusion amount corresponding to each deformed intrusion section, the initial boundary of the channel corresponding to the deformed intrusion section is corrected by shrinking inward, and spliced ​​with the initial boundary of the channel corresponding to the non-deformed intrusion section to obtain the effective passage boundary.

[0166] In one embodiment, the "travel envelope height range" referred to in this application refers to the height range within which the mobile robot body and its transported goods may actually interfere with facilities on both sides of the passage during the current working condition. For the mushroom transport scenario, the lateral interference height range of the robot under unloaded conditions is typically determined mainly by the lateral envelope of the vehicle body and the superstructure; while under transport conditions, the lifting or expansion effects of the carrier, pallet, cultivation bed frame, or other load-bearing components on the overall spatial envelope must also be considered. Therefore, the travel envelope height range can be determined based on the robot's own height dimensions, the transported goods' external dimensions, and their relative installation relationship. For example, in one embodiment, the lowest lateral interference height of the robot under transport conditions can be used as the lower bound, and the highest lateral interference height as the upper bound to form the travel envelope height range. This height range can be obtained directly from the equipment design parameters, or it can be determined through a configuration table or digital model using conventional shape modeling methods in the art.

[0167] In one implementation, the height range of the passage envelope should correspond to the current transport status. That is, the corresponding height range of the passage envelope may differ when the robot carries vehicles of different sizes or installation heights. For example, when using a standard mushroom transport vehicle, the height range of the passage envelope can be 0.25m to 1.35m above the ground; when only passing through a narrow passage empty, this range can be reduced to 0.15m to 0.95m above the ground. These values ​​are not given in isolation, but correspond to two typical working conditions: "overall passage envelope of the vehicle-mounted vehicle" and "passage envelope of the vehicle body." By introducing the height range of the passage envelope, subsequent boundary deformation judgments can focus on the height range that truly affects current passage, without over-processing changes in height layer boundaries unrelated to current passage.

[0168] In one embodiment, when determining the deformable intrusion section within the passage envelope height range, the current observed boundary of the mushroom cultivation rack within the height range can be obtained first, and then compared with the initial boundary of the passage to identify the boundary section that intrudes into the passage relative to the initial boundary. The "deformable intrusion section" referred to in this application refers to a boundary section in the passage extension direction where the facility boundary intrudes into the passage's passable area within a continuous segment, and this intrusion is sufficient to affect the robot's passage judgment. The corresponding "intrusion amount" refers to the offset of the current observed boundary within this section relative to the initial boundary of the passage into the passage. The intrusion amount can be represented by the maximum intrusion value, average intrusion value, or intrusion value at a representative sampling point within the section. To ensure that the boundary correction result balances safety margin and passage efficiency, in one embodiment, an intrusion amount characterization method that reflects the most unfavorable passage conditions of the section can be preferred, such as using the maximum intrusion value within the section, or using a combination of the maximum intrusion value and the average intrusion value of the section to form the correction basis.

[0169] In one embodiment, an intrusion determination threshold can be set when identifying deformable intrusion sections. This threshold can be determined based on the minimum safe passage margin, boundary measurement accuracy, and robot lateral control error. For example, the intrusion determination threshold can be set to 5mm to 25mm. When boundary measurement noise is low and the narrow channel passage margin is sensitive, 8mm or 10mm can be used; when boundary observation exhibits some fluctuation and it is necessary to avoid over-responding to small measurement fluctuations, 15mm or 20mm can be used. The above example values ​​correspond to high-sensitivity boundary intrusion identification scenarios and anti-measurement fluctuation scenarios, respectively. If the intrusion amount within a certain section exceeds the intrusion determination threshold, the section can be identified as a deformable intrusion section; otherwise, the initial boundary of the channel corresponding to that section can be preserved unchanged.

[0170] In one embodiment, when correcting the initial boundary of the passage by shrinking it inward based on the intrusion amount corresponding to each deformed intrusion segment, the corrected boundary position can be calculated separately for each deformed intrusion segment. Specifically, the boundary can be discretized into multiple segments along the passage extension direction. For each identified deformed intrusion segment, the initial boundary of the passage corresponding to that segment is corrected along the inner side of the passage based on its corresponding intrusion amount; for non-deformed intrusion segments, the original initial boundary of the passage remains unchanged. After completing the correction of each segment, the corrected boundaries of the deformed intrusion segments are continuously spliced ​​with the uncorrected boundaries of the non-deformed intrusion segments to obtain a complete and effective passage boundary. This segmented processing method avoids uniformly and conservatively shrinking the entire boundary, thereby preserving the passage capacity of the undeformed segments as much as possible while ensuring local safety margins.

[0171] In one embodiment, to ensure that the effective passage boundary after splicing is geometrically continuous and suitable for subsequent trajectory generation, a continuous transition process can be performed at the boundary between adjacent segments. For example, conventional curve smoothing or boundary transition methods in the art can be used to ensure that there are no obvious abrupt changes in the connection between the modified segment boundary and the unmodified segment boundary. In this way, on the one hand, the stability of candidate trajectory generation can be avoided due to overly discrete boundary representation, and on the other hand, the effective passage boundary can better conform to the continuous shape of the actual passable corridor.

[0172] For example, the above process can be illustrated using a mushroom transport scenario. For instance, when a mobile robot is transporting goods through a narrow passage outside the growth area, the vehicle it carries causes the robot's lateral travel envelope height to shift upwards. The system first determines the corresponding travel envelope height range based on the robot's current transport status, for example, 0.25m to 1.35m above the ground. Then, within this height range, it extracts the current observed boundary of the mushroom cultivation rack on one side of the passage and compares it with the initial boundary of the passage. If an inward intrusion boundary offset greater than 12mm is found within a continuous segment, this continuous segment can be identified as a deformed intrusion segment, and the initial boundary of the passage for this segment is corrected by shrinking it inwards based on its maximum intrusion amount; other segments that do not exceed the intrusion judgment threshold retain the original initial boundary of the passage. Subsequently, the corrected segment and the uncorrected segment are continuously spliced ​​together to form an effective travel boundary. In this way, the boundary conditions used for subsequent candidate trajectory generation and screening can accurately reflect the spatial range where boundary interference may actually occur under the current transport conditions.

[0173] It should be noted that the methods for determining the regional temperature and humidity transition parameters, calculating the regional environmental disturbance cost, the environmental sensitivity coefficient, the passage envelope height range, the intrusion judgment threshold, and the intrusion amount in the above embodiments can all be conventionally set by those skilled in the art in conjunction with the environmental control requirements of the mushroom workshop, robot size parameters, vehicle dimensions, and real-time control capabilities. As long as it still reflects the determination of the regional environmental disturbance cost based on the temperature and humidity transition situation in the regional type and candidate passage direction, and the consideration of this cost in the candidate trajectory sorting; and the identification of deformed intrusion sections based on the passage envelope height range corresponding to the carrying conditions and the correction of the initial channel boundary accordingly to obtain the effective passage boundary, it should be considered as the technical solution to be described in this application.

[0174] Based on the same inventive concept, this application also provides an obstacle avoidance trajectory generation system based on a diffusion strategy, which corresponds to an obstacle avoidance trajectory generation method based on a diffusion strategy. Since the principle of the system in this application is similar to the obstacle avoidance trajectory generation method based on a diffusion strategy described above, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be described again.

[0175] Reference Figure 4 The diagram shown is a schematic of an obstacle avoidance trajectory generation system based on a diffusion strategy provided in an embodiment of this application. The system includes:

[0176] The data acquisition module 10 is used to acquire the operating status information, task target information, and environmental status information of the mobile robot.

[0177] The first processing module 20 is used to determine the local planning area for local obstacle avoidance planning based on the environmental state information; determine the obstacle occupancy information within the local planning area; determine the boundary position and local deformation of the button mushroom cultivation rack within the local planning area; and determine the effective passage boundary based on the boundary position and the local deformation.

[0178] Encoding module 30 is used to encode the running status information, the operation target information, the obstacle occupancy information, and the effective passage boundary into guidance condition features;

[0179] The second processing module 40 is used to perform multi-step inverse diffusion denoising iteration based on multiple initial noise sequences and under the constraints of the guiding condition features to generate multiple candidate obstacle avoidance trajectories.

[0180] The selection module 50 is used to eliminate candidate obstacle avoidance trajectories that conflict with the obstacle occupancy information, exceed the effective passage boundary, or do not meet the execution constraints, and to determine the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for generating obstacle avoidance trajectories based on a diffusion strategy, characterized in that, include: Acquire information on the operating status, operational objectives, and environmental conditions of the mobile robot; Based on the environmental status information, determine the local planning area for local obstacle avoidance planning; determine the obstacle occupancy information within the local planning area; determine the boundary position and local deformation of the button mushroom cultivation rack within the local planning area, and determine the effective passage boundary based on the boundary position and the local deformation. The operating status information, the task target information, the obstacle occupancy information, and the valid passage boundary are encoded as guidance condition features; Starting from multiple initial noise sequences, a multi-step inverse diffusion denoising iteration is performed under the constraints of the guiding condition features to generate multiple candidate obstacle avoidance trajectories. Candidate obstacle avoidance trajectories that conflict with the obstacle occupancy information, exceed the effective passage boundary, or do not meet the execution constraints are eliminated, and the target obstacle avoidance trajectory is determined from the remaining candidate obstacle avoidance trajectories.

2. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 1, characterized in that, The determination of obstacle occupancy information within the local planning area includes: Based on the current motion state of the dynamic obstacle, predict the predicted position of the dynamic obstacle at each moment in the planning time domain; The obstacle occupancy information is determined based on the predicted location at each time point.

3. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 1, characterized in that, The determination of the valid passage boundary includes: Based on the design boundary of the button mushroom cultivation rack, determine the initial boundary of the passage; The initial boundary of the channel is corrected along the inner side of the channel based on the local deformation to obtain the effective passage boundary.

4. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 1, characterized in that, The multi-step inverse diffusion denoising iteration includes: During the inverse diffusion denoising iteration, the trajectory samples are guided and updated based on the obstacle occupancy information, the effective passage boundary, and the execution constraints.

5. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 1, characterized in that, Determining the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories includes: The remaining candidate obstacle avoidance trajectories are sorted according to the target positioning deviation and smoothness. The candidate obstacle avoidance trajectory with the best ranking result is determined as the target obstacle avoidance trajectory.

6. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 1, characterized in that, The step of determining the local planning area for local obstacle avoidance planning based on the environmental state information includes: The area where passage or arrival cannot be completed solely based on the predetermined driving trajectory under the current environmental conditions is identified as the local planning area.

7. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 1, characterized in that, Also includes: The mobile robot is controlled to execute the target obstacle avoidance trajectory in a rolling execution mode; During execution, in response to the current execution trajectory no longer meeting the passage conditions or the deviation of the current position from the current execution trajectory meeting the replanning conditions, the target obstacle avoidance trajectory is regenerated based on the updated current state.

8. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 5, characterized in that, Also includes: Determine the region type of the local planning area in the mushroom growth area, harvesting area, or buffer passage, and determine the regional temperature and humidity transition parameters of the local planning area along the candidate passage direction; Based on the cross-regional path, running speed, and regional temperature and humidity transition parameters corresponding to the candidate obstacle avoidance trajectory, the regional environmental disturbance cost is determined; the regional environmental disturbance cost is used to characterize the degree of influence of the candidate obstacle avoidance trajectory on the stability of the mushroom growth environment and the smoothness of the cross-regional environmental transition. When determining the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories, the regional environmental disturbance cost is used as one of the sorting criteria.

9. The obstacle avoidance trajectory generation method based on a diffusion strategy according to claim 3, characterized in that, The effective passage boundary is obtained by: Determine the passage envelope height range of the mobile operation robot under the carrying conditions; Within the height range of the passage envelope, the deformed intrusion sections of the mushroom cultivation rack that intrude into the inner side of the channel relative to the initial boundary of the channel, and the intrusion amount corresponding to each deformed intrusion section are determined; Based on the intrusion amount corresponding to each deformed intrusion segment, the initial boundary of the channel corresponding to the deformed intrusion segment is shrunken and corrected, and then spliced ​​with the initial boundary of the channel corresponding to the non-deformed intrusion segment to obtain the effective passage boundary.

10. A diffusion-based obstacle avoidance trajectory generation system, used to implement the diffusion-based obstacle avoidance trajectory generation method described in claims 1-9, characterized in that, include: The data acquisition module is used to acquire the mobile robot's operating status information, task target information, and environmental status information. The first processing module is used to determine the local planning area for local obstacle avoidance planning based on the environmental state information; determine the obstacle occupancy information within the local planning area; determine the boundary position and local deformation of the button mushroom cultivation rack within the local planning area; and determine the effective passage boundary based on the boundary position and the local deformation. The encoding module is used to encode the running status information, the operation target information, the obstacle occupancy information, and the effective passage boundary into guidance condition features; The second processing module is used to perform multi-step inverse diffusion denoising iteration based on multiple initial noise sequences and under the constraints of the guiding condition features to generate multiple candidate obstacle avoidance trajectories. The selection module is used to eliminate candidate obstacle avoidance trajectories that conflict with the obstacle occupancy information, exceed the effective passage boundary, or do not meet the execution constraints, and to determine the target obstacle avoidance trajectory from the remaining candidate obstacle avoidance trajectories.

Images