Mobile robot path optimization system and method supporting rail mode switching

Abstract

This invention belongs to the field of robot control, and particularly relates to a mobile robot path optimization system and method supporting track mode switching. The system uses an encoding module to integrate the mobile robot's operating scenario and real-time load status to generate a navigation task seed containing a path cost vector, navigation strategy identifier, and scenario transition point parameters. The central scheduler dynamically selects and calls the corresponding track navigation strategy sub-model or flat ground navigation strategy sub-model based on this navigation task seed. Combined with a spatiotemporal reservation table based on continuous value reservation weights to achieve multi-vehicle conflict avoidance and resource scheduling, executable trajectory control instructions are generated. This invention achieves load-adaptive path planning, efficient multi-vehicle collaboration, and smooth, seamless navigation switching between track and flat ground scenarios ensured by a switching prediction sub-model, significantly improving operational efficiency and system reliability in mixed and complex environments.

Description

Technical Field

[0001] This invention belongs to the field of robot control, and in particular relates to a mobile robot path optimization system and method that supports track mode switching. Background Technology

[0002] In modern warehousing, flexible manufacturing, and smart agriculture, mobile robots, such as four-way shuttles, are core equipment for achieving efficient and precise material flow. These robots often need to perform tasks in a mixed environment consisting of fixed track networks and open flat ground. In track areas, the robot operates at high speed and stably along structured paths; in flat ground areas, it needs autonomous obstacle avoidance and flexible path planning capabilities to cope with the dynamic environment. Current technologies for path planning and control in such mixed scenarios typically employ segmented processing or single-algorithm adaptation, but these methods have the following shortcomings: First, traditional path planning algorithms do not fully consider the impact of the robot's load state, a dynamic factor, on the actual travel cost; taking the widely used A... Taking the algorithm as an example, it searches for the optimal path through an evaluation function, which includes the actual cost from the starting point to the current node. However, in traditional implementations, the actual cost usually only includes static geometric or topological costs such as distance and number of turns, without incorporating load weight. In reality, the robot's load state directly affects its motion inertia, energy consumption, and braking performance. Especially in operation scenarios with frequent starts, stops, and turns, the acceleration energy and braking distance required for full-load operation are much greater than those for unloaded operation. Existing methods, by ignoring this factor, may result in low energy efficiency, slow response, or even safety issues due to insufficient braking distance estimation during actual execution of the planned optimal path. Secondly, single or fixed path planning strategies are difficult to adapt to dynamic and complex environments that mix tracks and flat ground. While the algorithm is efficient and ensures optimality on known, static track maps, it is inherently unsuitable for handling sudden dynamic obstacles or real-time changes in passage costs in flat terrain. It requires global replanning once the environment changes, resulting in poor real-time performance. Conversely, biomimetic algorithms such as ant colony optimization, through swarm intelligence and pheromone mechanisms, can better adapt to the dynamic and uncertain nature of the environment. However, their computational efficiency may be lower than that of A in highly structured track scenarios. Algorithms: Most existing systems deploy planning algorithms independently for two different scenarios, lacking a mechanism for smooth, intelligent algorithm switching during robot operation. This leads to discontinuous and abrupt changes in the robot's path when crossing the boundary between tracks and flat ground, requiring pauses or significant adjustments, thus hindering the overall smoothness and efficiency of the operation.

[0003] In summary, in scenarios where mobile robots perform transport tasks in mixed track and flat environments, how to design a control system that can dynamically respond to load changes to improve actual energy efficiency and safety, and achieve seamless and adaptive switching of planning algorithms between different scenarios to ensure overall operational efficiency, is a technical problem that needs to be solved in this field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a mobile robot path optimization system and method supporting track mode switching. This system, through an encoding module, integrates the mobile robot's operating scenario and real-time load status to generate a navigation task seed containing a path cost vector, navigation strategy identifiers, and scenario transition point parameters. The central scheduler dynamically selects and invokes the corresponding track navigation strategy sub-model or flatland navigation strategy sub-model based on this navigation task seed. Combined with a spatiotemporal reservation table based on continuous value reservation weights for multi-vehicle conflict avoidance and resource scheduling, executable trajectory control commands are generated. This invention achieves load-adaptive path planning, efficient multi-vehicle collaboration, and smooth, seamless navigation switching between track and flatland scenarios ensured by a switching prediction sub-model, significantly improving operational efficiency and system reliability in complex mixed environments.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A mobile robot path optimization system that supports track mode switching includes:

[0007] The encoding module comprehensively encodes the mobile robot's operating scenario and real-time load status to generate navigation task seeds;

[0008] The instruction generation module, based on the navigation task seed, selects the corresponding navigation strategy model and combines it with a preset spatiotemporal reservation table to generate control instructions for the executable trajectory, and outputs an instruction readiness indication to the central scheduler. The navigation strategy model includes a track navigation strategy sub-model, a flat ground navigation strategy sub-model, and a switching prediction sub-model. The spatiotemporal reservation table is used to calculate a continuous value reservation weight for each path node based on the three-dimensional coordinates or two-dimensional grid coordinates of each path node in the executable trajectory and its corresponding time window. This weight is associated with the probability of multi-vehicle conflict and the remaining safe passage time of the path node, so as to perform multi-vehicle collaborative conflict avoidance and resource scheduling for the selected track navigation strategy sub-model or flat ground navigation strategy sub-model.

[0009] Upon receiving the instruction ready indication, the control module sends the control instruction for the executable trajectory to the corresponding track or flat ground navigation control loop.

[0010] Specifically, the orbital navigation strategy sub-model is composed of improved A. The algorithm is constructed by combining an improved APF and is used to calculate the local optimal trajectory in a structured track scenario. The flat ground navigation strategy sub-model is constructed by combining the ACO algorithm with APF and is used to calculate the local optimal flat ground path trajectory in a flat ground scenario. The switching prediction sub-model is used to calculate the optimal time point for navigation strategy switching and the remaining time length for navigation strategy switching based on the real-time position coordinates of the mobile robot in the corresponding scenario and the scene switching point parameters in the pre-stored map data.

[0011] Specifically, the navigation task seed includes: a path cost vector, which at least includes the actual distance from the starting point to the current estimated node, the estimated number of turns, and a first inertial force influence coefficient calculated based on the real-time load state of the mobile robot; a navigation strategy identifier, used to indicate the currently selected navigation strategy model, whose value range includes at least a first identifier value for the track scene and a second identifier value for the flat ground scene; and scene transition point parameters, which include the three-dimensional coordinate position at the boundary between the track scene and the flat ground scene obtained from the pre-stored map data, and the estimated remaining time to reach the boundary calculated in real time based on the current speed of the mobile robot and the three-dimensional coordinate position.

[0012] Specifically, the construction process of the orbital navigation strategy sub-model includes:

[0013] Obtain the current weight of the robot's load, structured environment map data, and the coordinates of the starting point and ending point of the local path planning.

[0014] Based on the current weight of the cargo loaded on the robot, and combined with the predefined full-load reference inertial force, the first inertial force influence coefficient is calculated.

[0015] Initialize a configured list of open nodes and a list of closed nodes; encapsulate the starting coordinates of the local path planning of the track into an initial path node, the attributes of which include at least node coordinates, cumulative actual cost, and heuristic cost; set the cumulative actual cost of the initial path node to an initial value and add it to the list of open nodes;

[0016] The node with the smallest comprehensive evaluation function value is selected from the list of open nodes as the current node to be expanded; for each traversable adjacent node of the current node to be expanded, a temporary cumulative actual cost is calculated; the temporary cumulative actual cost is constructed from at least the base distance cost, the turning action cost, the elevation action cost, the first inertial force influence coefficient, the preset edge weight penalty value, and the dynamic reservation weight value determined by querying the spatiotemporal reservation table.

[0017] Specifically, the construction process of the orbital navigation strategy sub-model also includes:

[0018] For each of the accessible adjacent nodes, after calculating the temporary cumulative actual cost, the following logical judgment and update operations are performed:

[0019] Determine whether the passable adjacent node has been recorded in the closed node list or the open node list;

[0020] If the accessible adjacent node is not recorded, a node record is created for it, its cumulative actual cost is set to the temporary cumulative actual cost, its parent node pointer is pointed to the current node to be expanded, and the improved heuristic function value of the current accessible adjacent node is calculated based on the node coordinates of the current accessible adjacent node, the global target point coordinates, and the environmental potential field constructed by the sensor or map. The node record of the current accessible adjacent node is then added to the list of open nodes.

[0021] If the currently accessible neighboring node already exists in the open node list or the closed node list, then the temporary cumulative actual cost is compared with the cumulative actual cost already recorded for the currently accessible neighboring node; if the temporary cumulative actual cost is less than the recorded cumulative actual cost, then the cumulative actual cost of the currently accessible neighboring node is updated to the temporary cumulative actual cost, and its parent node pointer is pointed to the currently expandable node; if the currently accessible neighboring node was originally in the closed node list, then it is moved back to the open node list.

[0022] Specifically, the construction process of the orbital navigation strategy sub-model also includes:

[0023] After completing the expansion process for all accessible adjacent nodes of the current node to be expanded, the current node to be expanded is moved from the open node list to the closed node list;

[0024] Repeat the steps of selecting the current node to be expanded from the list of open nodes, calculating the temporary cumulative actual cost and improved heuristic function value of the passable adjacent nodes, comparing and updating the cost, and transitioning the node state until any of the following termination conditions are met: the node corresponding to the endpoint coordinate of the local path planning of the track is added to the list of closed nodes, or the list of open nodes is empty;

[0025] When the node corresponding to the endpoint coordinate of the local path planning is added to the list of closed nodes, starting from the endpoint node coordinate of the local path planning, the path is traced backward along its parent node pointer to the initial path node, generating a locally optimal track path trajectory from the starting coordinate of the local path planning to the endpoint coordinate of the local path planning.

[0026] Specifically, the flat terrain navigation strategy sub-model is used to calculate the locally optimal flat terrain path trajectory in a flat terrain scenario, including:

[0027] Acquire 2D grid map data of flat terrain scene, coordinates of the starting point and ending point of local path planning on flat terrain, and obstacle information;

[0028] For each grid node in the map, calculate a gravitational potential field value based on the distance from the grid node to the target point, and a repulsive potential field value based on the distance from the grid node to the nearest obstacle;

[0029] The core parameters for initializing the ant colony optimization algorithm include the preset ant population size, maximum number of iterations, basic pheromone volatility coefficient, pheromone enhancement constant, first heuristic factor characterizing pheromone importance, and second heuristic factor characterizing distance heuristic importance.

[0030] An initial pheromone concentration matrix is ​​initialized. Based on the starting coordinates of the local path planning on flat ground, the ending coordinates of the local path planning on flat ground, and the gravitational potential field value, an initial pheromone concentration is assigned to each node. The initial pheromone concentration is positively correlated with the gravitational potential field value of the corresponding grid node and negatively correlated with the ending information of the previous adjacent local optimal orbit path trajectory.

[0031] Specifically, the flat terrain navigation strategy sub-model, used to calculate the locally optimal flat terrain path trajectory in a flat terrain scenario, also includes:

[0032] In each iteration, a path search is performed independently for each virtual ant; the path search is conducted in a network composed of the two-dimensional raster map data of the flat terrain scene, where each raster node is regarded as a path node, and the path search is specifically performed as follows:

[0033] Based on the ant's current path node, calculate the state transition probability that it can move to the next candidate path node; the state transition probability includes the pheromone concentration corresponding to the candidate path node, the heuristic value based on the distance between path nodes, and the repulsive potential field value of the candidate path node, wherein the repulsive potential field value is used as a penalty factor to reduce the probability of grid nodes near obstacles being selected.

[0034] Based on the calculated state transition probabilities of each candidate path node, the roulette wheel selection method is used to determine the next moving path node of the ant, and the selected path node is added to the current ant's current path sequence.

[0035] Repeat the state transition and path node selection process until the ant reaches the local path planning endpoint coordinates on flat ground or the path construction fails.

[0036] Specifically, the flat terrain navigation strategy sub-model, used to calculate the locally optimal flat terrain path trajectory in a flat terrain scenario, also includes:

[0037] Repeat the iterative process consisting of the ant path iterative construction step and the pheromone global update and collaborative adjustment step until the preset maximum number of iterations is reached or the pheromone distribution converges, specifically:

[0038] The steps for iteratively constructing an ant path include: independently performing a complete path search from the starting point to the ending point for each virtual ant in the ant colony;

[0039] The steps for global pheromone update and collaborative adjustment include: updating the pheromone concentration of the entire graph according to predefined rules based on the paths constructed by all ants in the current iteration and the dynamic reservation weight values ​​obtained from the spatiotemporal reservation table;

[0040] From all the paths generated in all iterations, the path with the highest comprehensive evaluation is selected as the optimal solution for this round of planning; the comprehensive evaluation of the path is determined based on path length, smoothness index, and safety index.

[0041] The optimal decoding is converted into an ordered sequence of path points consisting of a series of continuous two-dimensional coordinate points, which is output as the local optimal flat terrain path trajectory.

[0042] Specifically, the process of building the switching prediction sub-model includes:

[0043] Based on the real-time position coordinates, real-time running speed of the mobile robot, and the coordinates of the transition connection point included in the scene transition point parameters, the estimated remaining time from the real-time position coordinates to the transition connection point coordinates is calculated.

[0044] Based on a preset switching delay and safety margin, a switching lead threshold is set; when the estimated remaining time is equal to or less than the switching lead threshold, the current moment is determined to be the optimal trigger time for navigation strategy switching.

[0045] Before reaching the optimal trigger time, initiate the cross-scenario artificial potential field data conversion process, specifically as follows:

[0046] When the predicted next target scene is a flat scene, a two-dimensional artificial potential field data pre-generation operation is performed, including: acquiring the three-dimensional artificial potential field data constructed under the current orbit scene, using a preset oblique projection formula to map the three-dimensional repulsive field values ​​in the three-dimensional artificial potential field data to the two-dimensional plane corresponding to the flat scene, and generating an initial repulsive field dataset for the flat scene suitable for two-dimensional raster maps by assigning corresponding projection weights to the three-dimensional repulsive field values ​​of different height layers and performing vector synthesis.

[0047] Specifically, the process of constructing the switching prediction sub-model also includes:

[0048] The coordination and initialization of navigation model parameters are performed in parallel with the cross-scene artificial potential field data conversion step, specifically including:

[0049] When the predicted next target scene is a flat scene, the robot receives the first path geometric information from the robot's current position to the coordinates of the transition connection point, output by the current track navigation strategy sub-model.

[0050] Based on the first path geometry information, the pheromone concentration matrix of the ant colony optimization algorithm in the flat ground navigation strategy sub-model is initialized in a biased manner, so that the grid nodes located in the direction indicated by the first path geometry information obtain the updated initial pheromone concentration.

[0051] Specifically, the process of constructing the switching prediction sub-model also includes:

[0052] When the predicted next target scenario is a track scenario, the path optimization feedback data output by the current flat terrain navigation strategy sub-model is received; the path optimization feedback data includes at least the key node sequence or curvature distribution information that characterizes the geometric smoothness of the path, generated during the iterative solution of the flat terrain scenario.

[0053] Based on the path optimization feedback data, adjust the improved A in the orbital navigation strategy sub-model. The gravitational potential adjustment coefficient corresponding to the potential field heuristic term on which the improved algorithm heuristic function depends;

[0054] When the mobile robot actually runs to the coordinates of the transition point, the central scheduler switches the currently active navigation strategy sub-model to the target navigation strategy sub-model that has completed data and parameter preloading according to the navigation strategy identifier. The target navigation strategy sub-model uses the pre-transformed initial potential field dataset and pre-loaded optimization parameters to generate continuous control commands for the new scenario.

[0055] Mobile robot path optimization methods that support track mode switching include:

[0056] The operating scenario and real-time load status of the mobile robot are comprehensively encoded to generate navigation task seeds;

[0057] Based on the navigation task seed, the corresponding navigation strategy model is selected and combined with the preset spatiotemporal reservation table to generate control instructions for the executable trajectory, and an instruction readiness indication is output to the central scheduler. The navigation strategy model includes a track navigation strategy sub-model, a flat ground navigation strategy sub-model, and a switching prediction sub-model. The spatiotemporal reservation table is used to calculate a continuous value reservation weight associated with the probability of multi-vehicle conflict and the remaining safe passage time of the path node for each path node based on the three-dimensional coordinates or two-dimensional grid coordinates of each path node in the executable trajectory and its corresponding time window, so as to optimize the selected track navigation strategy sub-model or flat ground navigation strategy sub-model for multi-vehicle collaborative conflict avoidance and resource scheduling.

[0058] Upon receiving the instruction ready indication, the control instruction for the executable trajectory is sent to the corresponding track or flat ground navigation control loop.

[0059] Compared with the prior art, the beneficial effects of the present invention are:

[0060] This invention addresses the shortcomings of existing technologies by using an encoding module to comprehensively encode the mobile robot's operating scenario and real-time load status to generate navigation task seeds. This allows for a close correlation between navigation strategy selection and path planning and load variation characteristics. By combining the inertial force influence coefficient calculated from the load, path costs can be accurately matched, effectively improving the actual energy efficiency during navigation. Simultaneously, based on the navigation task seeds, a navigation strategy model is selected that includes track, flat ground, and switching prediction sub-models. Furthermore, the reserved weights of path nodes are calculated using a spatiotemporal reservation table, leveraging an improved A... The sub-model, which integrates the algorithm, ACO algorithm, and APF, realizes local optimal path planning in both structured track and unstructured flat terrain scenarios. It also accurately calculates the optimal time point and remaining time by switching the prediction sub-model. Combined with the optimization of multi-vehicle conflict probability and node reservation time weight in the spatiotemporal reservation table, it achieves seamless and adaptive switching of navigation algorithms in different scenarios. Furthermore, it can improve the safety and global operating efficiency of multi-vehicle collaborative operation through conflict avoidance and resource scheduling. Attached Figure Description

[0061] Figure 1 This is a block diagram of a mobile robot path optimization system supporting track mode switching according to Embodiment 1 of the present invention;

[0062] Figure 2 This is a flowchart of the mobile robot path optimization method supporting track mode switching in Embodiment 2 of the present invention. Detailed Implementation

[0063] Example 1:

[0064] To facilitate understanding of the background and scope of this application, the following description of specific implementation methods is provided in the context of a multi-level track-flat ground hybrid operation scenario in a smart agricultural greenhouse.

[0065] In large-scale greenhouse planting bases for smart agriculture, four-way vehicles serve as the core carrier for automated material transfer. They are responsible for the entire process of transferring mature fruits and vegetables from the track storage location in the planting area to the sorting area on the flat ground, and empty boxes from the supply area on the flat ground to the track in the planting area. The operation scenario presents complex characteristics of three-dimensional track and two-dimensional flat ground intertwining, dense multi-vehicle collaboration, and dynamic load changes.

[0066] Within the same work cycle, multiple four-way vehicles need to shuttle between three levels of tracks, while frequently switching to flat ground to complete loading, unloading, and sorting connections. The track area has a multi-level, multi-lane intersection structure, with static obstacles such as curves and support pillars scattered throughout, while the flat ground area faces dynamic obstacles such as workers, temporary material boxes, and sorting equipment. In practice, the four-way vehicles need to adapt to dynamically changing load states to perform transfer tasks, and task requirements are dynamically adjusted according to the harvesting progress. For example, during peak harvesting periods, the transfer frequency needs to be increased, while resources such as track node access rights and flat ground operating space are subject to strict time window constraints. The aforementioned complex characteristics are amplified by a common phenomenon: multiple four-way vehicles are prone to spatiotemporal conflicts in key areas such as cross-level track intersections and track-flat ground connection points. At the same time, the inertial force under full load can easily cause cargo to sway or even overturn during sharp turns, and the navigation strategy discontinuity during scene switching can easily cause sudden trajectory changes and operational stoppages.

[0067] Existing four-way vehicle path planning methods mostly employ two routes: one is to separate the track from the flat ground scene, with the track end using traditional A / B algorithm. The algorithm plans the path separately for flat terrain using the Ant Colony Algorithm (ACO), and directly triggers the algorithm replacement when the scene changes. Secondly, it ignores the impact of dynamic load changes and adopts a unified path planning model with a fixed cost coefficient, and achieves obstacle avoidance through subsequent collision detection and path adjustment.

[0068] When switching between track and flat terrain, the former method, due to the lack of prior potential field data conversion and parameter coordination, is prone to drastic changes in path direction and interruptions in control commands. This can cause four-way vehicles to decelerate and stop at connection points, or even replan their routes, significantly reducing transfer efficiency. Furthermore, traditional algorithms do not consider the inertial forces associated with the load, making them prone to planning sharp turns when fully loaded, increasing the risk of cargo swaying. While the latter method simplifies model complexity, its fixed cost coefficient cannot adapt to the differences in inertial forces under different loads. Moreover, its collision detection-based obstacle avoidance is a passive response, which can lead to untimely obstacle avoidance and repeated path adjustments in scenarios with dense multi-vehicle coordination, potentially causing multi-vehicle congestion and failing to meet the efficiency and safety requirements of greenhouse transport.

[0069] The common challenges of the two routes mentioned above are: first, the dynamic changes in load and the geometric constraints of the scene, such as the coupling effect of track curves, are not incorporated into the path cost calculation, resulting in the planned path failing to match the actual operational stability requirements; second, the lack of a forward-looking coordination mechanism for switching navigation strategies between track and flat terrain scenarios leads to a disconnect between perception and control; and third, multi-vehicle collaborative obstacle avoidance relies on post-detection and adjustment, and the lack of a forward-looking mechanism for predicting and avoiding spatiotemporal conflicts makes it difficult to adapt to the dynamic operational needs of complex mixed scenarios.

[0070] The core processing logic of this application is based on the aforementioned common dilemma. It no longer isolates the navigation problems of different scenarios by separating load states and scene characteristics. Instead, it first identifies the core contradiction in smart agriculture four-way vehicle operation as the synergistic satisfaction of operational stability under dynamic loads, smooth continuity of cross-scene navigation, and spatiotemporal safety of multi-vehicle collaboration, and prioritizes these three as the core constraints of path planning. To this end, after acquiring a greenhouse 3D track-ground fusion map and the real-time status of the four-way vehicles, including position, speed, load, and task requirements, an integrated framework for encoding, strategy selection, and collaborative control is constructed.

[0071] The coding module integrates the operational scenario and real-time load status into a navigation task seed, clearly defining the navigation strategy identifier, path cost vector, and scenario transition point parameters, thus providing a foundation for differentiated strategy selection. Improved A / B models are constructed to address the differences in characteristics between track and flat terrain scenarios. This paper presents two sub-models: a track navigation strategy combining the algorithm and APF, and a flat-ground navigation strategy combining the ACO algorithm and APF. A switching prediction sub-model is also designed to achieve cross-scenario potential field data pre-conversion and parameter pre-initialization. A spatiotemporal reservation table maintained by a central scheduler is introduced to quantify the risk of multi-vehicle spatiotemporal conflicts into dynamic reservation weight values, integrating path cost calculation upfront to achieve proactive collision avoidance. Unlike conventional planning methods that fragment scenarios, ignore load, and passively avoid obstacles, this embodiment integrates load adaptation, scenario collaboration, and multi-vehicle prediction throughout the entire path planning process. Through iterative parameter optimization, the overall performance of the path is continuously improved, forming a full-link path optimization scheme adapted to the mixed scenarios of smart agriculture.

[0072] It should be noted that, given the critical need for stability and efficiency in greenhouse operations, this application employs an adaptation logic of precise load sensing, dynamic cost calculation, and iterative parameter optimization.

[0073] The inertial forces of the four-way vehicle vary significantly under different load conditions, and a simple fixed-parameter model cannot simultaneously meet the operational requirements of both empty and fully loaded conditions. Furthermore, the obstacle distribution in a greenhouse setting, such as track support columns and temporary material bins on flat ground, exhibits dynamic changes, making it difficult for a fixed path cost coefficient to adapt to environmental variations. This embodiment accurately acquires the load status through weighing sensors and multi-sensor fusion, calculates the inertial force influence coefficient, and dynamically adjusts the path cost. Simultaneously, based on historical operational data, a comprehensive performance score and parameter loss function are constructed, and model parameters are iteratively optimized using algorithms such as gradient descent. This ensures that the planned path approaches global optimality in terms of length, time, energy consumption, and safety, avoiding the insufficient adaptability issues caused by fixed parameters.

[0074] It should be emphasized that the above description is not an exhaustive list of algorithmic details, but rather clarifies the technical starting point and reproducible implementation boundaries of this embodiment:

[0075] With a known 3D track-ground integrated map of the greenhouse, real-time data collection of the four-way vehicle status and task requirements, and unified coding of constraint rules such as safety distance and operating speed limits, precise path planning for cross-scene and multi-vehicle collaboration under dynamic loads is achieved through scene-load integrated coding, differentiated navigation strategy sub-model construction, cross-scene forward-looking switching, and multi-vehicle spatiotemporal collaborative prediction mechanisms. This approach pre-processes the coupling effects of load and scene, transforms the perception-control gap in cross-scene switching into forward-looking collaboration, and converts the spatiotemporal conflict risks of multiple vehicles into quantifiable path costs. In principle, it resolves the triple challenges of operational stability, cross-scene smoothness, and multi-vehicle collaborative safety in existing technologies, providing a unified technical framework for the efficient and intelligent upgrading of automated material transportation in smart agricultural greenhouses.

[0076] Please see Figure 1 One embodiment of the present invention provides a mobile robot path optimization system supporting track mode switching, comprising:

[0077] The encoding module comprehensively encodes the mobile robot's operating scenario and real-time load status to generate a navigation task seed. The navigation task seed specifically includes: a path cost vector, which at least contains the actual distance from the starting point to the current estimated node, the estimated number of turns, and a first inertial force influence coefficient calculated based on the mobile robot's real-time load status; a navigation strategy identifier, used to indicate the navigation strategy model to be selected, whose value range includes at least a first identifier value for the track scenario and a second identifier value for the flat ground scenario; and a scene transition point parameter, containing the three-dimensional coordinate position at the boundary between the track scenario and the flat ground scenario obtained from pre-stored map data, and the estimated remaining time to reach the boundary calculated in real-time based on the mobile robot's current speed and the three-dimensional coordinate position. It should be further explained that the operational scenario data in this embodiment is acquired through laser obstacle avoidance sensors, environmental perception sensors, and pre-stored map data. It identifies in real time whether the current scenario is a structured track scenario or an unstructured flat ground scenario, and simultaneously collects environmental information such as path node coordinates, obstacle distribution, and scenario boundaries. Real-time load status data is collected by weighing and force sensors integrated at the fork root, lifting mechanism, or drive wheel axle, such as strain gauge or piezoelectric sensors, to collect the weight of the cargo. This data is combined with an inertial measurement unit to obtain the vehicle body attitude, and combined with parameters such as motor armature current and speed collected by Hall sensors. After processing by filtering algorithms, the load weight, center of gravity distribution, and inertial-related data are accurately obtained. Both types of data support real-time acquisition and transmission, providing a foundation for subsequent coding and navigation strategy selection.

[0078] The instruction generation module, based on the navigation task seed, selects the corresponding navigation strategy model and combines it with a preset spatiotemporal reservation table to generate control instructions for the executable trajectory, and outputs an instruction readiness indication to the central scheduler. The navigation strategy model includes a track navigation strategy sub-model, a flat ground navigation strategy sub-model, and a switching prediction sub-model. The spatiotemporal reservation table is used to calculate a continuous value reservation weight for each path node based on the three-dimensional coordinates or two-dimensional grid coordinates of each path node in the executable trajectory and its corresponding time window. This weight is associated with the probability of multi-vehicle conflict and the remaining safe passage time of the path node, so as to perform multi-vehicle collaborative conflict avoidance and resource scheduling for the selected track navigation strategy sub-model or flat ground navigation strategy sub-model.

[0079] It should be further noted that the orbital navigation strategy sub-model in this embodiment is derived from the improved A model. The algorithm is constructed by combining an improved APF and is used to calculate the local optimal trajectory in a structured track scenario. The flat ground navigation strategy sub-model is constructed by combining the ACO algorithm with APF and is used to calculate the local optimal flat ground path trajectory in a flat ground scenario. The switching prediction sub-model is used to calculate the optimal time point for navigation strategy switching and the remaining time length for navigation strategy switching based on the real-time position coordinates of the mobile robot in the corresponding scenario and the scene switching point parameters in the pre-stored map data.

[0080] Upon receiving the instruction ready indication, the control module sends the control instruction for the executable trajectory to the corresponding track or flat ground navigation control loop.

[0081] It should be further explained that the construction process of the orbital navigation strategy sub-model in this embodiment includes:

[0082] A101. Obtain the current weight of the robot's load, structured environment map data, and the starting and ending coordinates of the local path planning. For example, in a three-layer track system of a smart agricultural greenhouse, including layers Z1, Z2, and Z3, the starting coordinates of the local path planning specifically represent the precise three-dimensional position of the four-way vehicle when fully loaded with 100kg tomato boxes. For example, the track stopping node at column Y8 of aisle X3 in layer Z2 can be represented as (X3, Y8, Z2), which is the starting reference point for path planning. The ending coordinates of the local path planning specifically represent the precise three-dimensional position of the target that the four-way vehicle needs to reach. For example, the target cargo position at column Y15 of aisle X5 in layer Z1 can be represented as (X5, Y15, Z1), which is the ending target point for path planning. Both together constitute the improved A101. The algorithm, combined with the APF-based orbital navigation strategy sub-model, provides the core boundary parameters for path search, which, in conjunction with the system's dynamically adjusted APF for full-load conditions, enables further optimization. The algorithm's cost-effectiveness enables the calculation of locally optimal trajectory paths by reducing the number of turns from 5 to 3.

[0083] A102. Based on the current weight of the cargo loaded on the robot, and combined with the predefined full-load reference inertial force, the first inertial force influence coefficient is calculated. Specifically: ,in W Given the current weight of the cargo carried by the robot, The preset full-load reference inertial force for the robot, This is a proportionality coefficient, set by those skilled in the art;

[0084] A103. Initialize a configured list of open nodes and a list of closed nodes; encapsulate the starting coordinates of the local path planning of the track into an initial path node, the attributes of which include at least node coordinates, cumulative actual cost, and heuristic cost; set the cumulative actual cost of the initial path node to an initial value and add it to the list of open nodes;

[0085] A104. Select the node with the smallest comprehensive evaluation function value from the list of open nodes as the current node to be expanded; for each traversable adjacent node of the current node to be expanded, calculate a temporary cumulative actual cost; the temporary cumulative actual cost is constructed from at least the base distance cost, turning action cost, elevation action cost, the first inertial force influence coefficient, the preset edge weight penalty value, and the dynamic reserved weight value determined by querying the spatiotemporal reserved table; in this embodiment, under the application background of four-way vehicle track and flat ground hybrid navigation in smart agriculture scenario, the base distance cost is obtained by calculating the Euclidean distance between path nodes, where the track scenario is a three-dimensional coordinate distance and the flat ground scenario is a two-dimensional grid distance, and the turning action cost is based on the distance from the starting point to the edge. The cost of lifting and lowering actions is calculated by combining the number of turns at the current node, the angle difference between the current direction of the four-way vehicle and the direction pointed to by the target node, and the first inertial force influence coefficient corresponding to the weight of the load. The cost of lifting and lowering actions is calculated by the number of lifting and lowering actions from the starting point to the current node, the height change difference, the energy consumption coefficient of the lifting mechanism, and the inertial loss coefficient associated with the weight of the load. The construction of the temporary cumulative actual cost in this embodiment also includes a low-speed travel time cost. This cost is calculated by multiplying the estimated travel time of the mobile robot between path nodes at low speed with the energy consumption coefficient corresponding to low-speed operation. It is used to optimize the time and energy consumption cost assessment of the expected low-speed travel segment in path planning and, together with other cost items, achieves global multi-objective optimal path search. The core motivation for introducing the low-speed travel time cost in this embodiment is to resolve the potential contradiction between the geometric shortest path and the time / energy consumption optimal path. Its basic principle is to explicitly model the spatiotemporal and energy consumption costs of the speed strategy. Specifically, this cost term estimates the time the mobile robot needs to travel at low speeds to meet safety constraints in specific road sections, such as curves, intersections, and near-target areas. This time is then multiplied by a specific energy consumption coefficient under low-speed conditions, quantifying the additional time and continuous energy costs incurred by the proactive deceleration control strategy into a calculable path cost. Its fundamental function is to force the path search algorithm to consider not only the geometric length and motion overhead of the path during planning, but also to proactively assess the impact of the speed curves implied by different path options on the total task duration and total energy consumption. This allows the algorithm to proactively avoid road sections that, while having short geometric distances, require prolonged low-speed crawling, ultimately achieving global multi-objective collaborative optimization of time, energy consumption, and safety, all while ensuring safety.

[0086] The preset edge weight penalty value in this embodiment is mainly constructed from the following variables: basic edge weight, path node repulsion value, and maximum repulsion threshold. The basic edge weight is a preset baseline cost; the path node repulsion value is calculated using a modified artificial potential field method, specifically representing the repulsion potential field strength of the current path node affected by surrounding obstacles; the maximum repulsion threshold is a preset normalization parameter. The calculation of the edge weight penalty value is based on the basic edge weight and dynamically increased according to the ratio of the path node repulsion value to the maximum repulsion threshold. This embodiment dynamically transforms the path node repulsion value, which reflects the real-time obstacle threat, into the additional passage cost of the corresponding edge in the path search graph, thereby guiding the algorithm to actively avoid obstacle areas at the global path planning level and generating a safer movement trajectory. For example, in the smart agricultural greenhouse scenario of this embodiment, the track system has a multi-layered, multi-lane structure. When a four-way vehicle fully loaded with 100kg tomato crates travels to the curve area of ​​the track, the modified artificial potential field method will sense the obstacles formed by the geometric constraints of the curve in real time and calculate a higher path node repulsion value for the path nodes on the inside of the curve. According to the mechanism of this embodiment, after comparing the path node repulsion value with the preset maximum repulsion threshold, it is dynamically converted into a significantly increased edge weight penalty value and superimposed on the improved A In the path cost calculation of the algorithm, its direct technical effect is that, during the global path planning stage, the algorithm actively abandons paths that may lead to sharp turns and are close to the inside of the curve due to the high penalty cost, and instead selects a smoother track with a larger turning radius. This achieves the specific application effect of avoiding sharp turns at track curves and reducing the impact of inertial forces through APF modification, while ensuring the safety of fully loaded transportation.

[0087] It should be further explained that the modified artificial potential field method in this embodiment is used to quantify the repulsion strength of path nodes affected by surrounding obstacles in the global path planning stage, and dynamically convert this strength into additional passage cost in the path search graph to guide the algorithm to actively avoid obstacles.

[0088] In this embodiment, the virtual resultant force direction constraint acting on the path node is calculated using the modified artificial potential field method. This resultant force is the vector sum of attraction and repulsion.

[0089] The calculation of the repulsive force is defined by the modified repulsive field formula: when the path node i The distance d between the object b and the obstacle b i , b Less than the preset impact threshold At that time, the magnitude of the repulsive force is related to The distance gradient is proportional to the square of the reciprocal of d(i,b), otherwise it is 0.

[0090] In this embodiment, the dynamic reservation weight value determined by querying the spatiotemporal reservation table is as follows:

[0091] Based on the current operating scenario type of the mobile robot, the specific coordinate information of the passable adjacent nodes, and the timestamp of the mobile robot's expected arrival at the node, a formatted query instruction is generated; wherein, if the current operating scenario is determined to be a track scenario, the specific coordinate information adopts a three-dimensional coordinate format; if the current operating scenario is determined to be a flat ground scenario, the specific coordinate information adopts a two-dimensional raster coordinate format.

[0092] The formatted query command is sent to the spacetime reservation table maintained uniformly by the central scheduler;

[0093] Receive the query response returned by the spatiotemporal reservation table, wherein the query response encapsulates a dynamic reservation weight value with continuous numerical values;

[0094] The dynamic reservation weight value is generated in real time by the spatiotemporal reservation table according to the following calculation logic: First, the repulsive potential energy values ​​generated around the query node by all mobile robots already on the planned path are aggregated to obtain an aggregated value, and the probability of conflict at the expected arrival time of the query node is estimated based on this aggregated value; second, combined with the time window information of the node being reserved, the remaining safe passage time from the expected arrival time is calculated; finally, the spatiotemporal reservation table calculates the dynamic reservation weight value through an internal function. This function ensures that the calculated dynamic reservation weight value is positively correlated with the estimated probability of conflict and negatively correlated with the remaining safe passage time, thereby accurately quantifying the passage conflict risk level and time urgency of the path node at a specific future moment; wherein the dynamic reservation weight value The specific formula is as follows: , ,in The preset weighting weights, The probability of a conflict occurring at path node i represents the probability of all four-way vehicles colliding at path node i, where N is the number of all four-way vehicles at path node i; the time variable t is defined as the estimated time when the mobile robot plans to arrive at path node i, which is expressed by the formula... w ( i,t )and T r ( i,t The key input parameters ensure that the calculation of the dynamic reserved weight value is based on the node conflict risk and time urgency at a specific future time t, thus making the path planning forward-looking and able to assess the risk level faced if the path node i is reached at time t. This is the core basis for achieving precise spatiotemporal coordinated scheduling. For the first sThe APF repulsive potential energy value generated by a four-way vehicle at path node i; The maximum repulsive force threshold is a normalization constant used to scale the sum of the repulsive potential energy values ​​to between 0 and 1, and is set by those skilled in the art. Let be the remaining safe passage time for path node i, when a node is about to be occupied. It will tend towards zero, and the weight will become infinitely large, thus having a strong deterrent effect; and The preset weighting values ​​are set by those skilled in the art based on historical values. It should be further explained that the core motivation for designing this formula in this embodiment is to enable the mobile robot's path planning to proactively perceive and quantify the risk of future spatial-temporal conflicts and time urgency. The principle is to fuse dynamic information from two dimensions: first, aggregating the repulsive potential energy generated by all other robots at the target node to estimate the probability of a collision or congestion at that location; second, querying the schedule of the node's reserved space to calculate the remaining safe passage time window from the expected arrival time. Finally, the function weights and fuses the conflict probability with the reciprocal of the remaining safe passage time, allowing the dynamic weight value to simultaneously respond to spatial conflict risk and time urgency. When the node conflict risk is high or the available time is extremely short, the weight value increases sharply, strongly guiding the algorithm to avoid the node in path cost calculation, thus achieving safe and efficient spatial-temporal collision avoidance under multi-robot collaboration.

[0095] A105. For each of the aforementioned passable adjacent nodes, after calculating the temporary cumulative actual cost, perform the following logical judgment and update operations:

[0096] A1051. Determine whether the passable adjacent node has been recorded in the closed node list or the open node list;

[0097] A1052. If the accessible adjacent node is not recorded, a node record is created for it, its cumulative actual cost is set to the temporary cumulative actual cost, its parent node pointer is pointed to the current node to be expanded, and based on the node coordinates of the current accessible adjacent node, the global target point coordinates, and the environmental potential field constructed by the sensor or map, the improved heuristic function value of the current accessible adjacent node is calculated, and the node record of the current accessible adjacent node is added to the open node list.

[0098] For example, suppose the current node to be expanded is a, and its three-dimensional coordinates are (X_a, Y_a, Z_a). It finds an eastward accessible adjacent node m with coordinates (X_a+1, Y_a, Z_a).

[0099] If node m is not recorded in any list, a new node record is created for m, including:

[0100] Set its cumulative actual cost g(m) to the calculated temporary cumulative actual cost g_temp;

[0101] Set its parent node pointer to point to node a;

[0102] Based on the coordinates of node m, the coordinates of the global target point G, and the environmental potential field, the improved heuristic function value f(m) = g(m) + h(m) + P(m) is calculated. Here, h(m) is the geometric heuristic value from m to G, and P(m) is the potential field heuristic term calculated based on the potential field at node m.

[0103] Add the node record of the accessible adjacent node m containing the above attributes to the open node list for future expansion.

[0104] This process adds node m as a candidate point for a new path from the starting point through node a to the search space, and its cost and heuristic information will guide the subsequent search direction.

[0105] A1053. If the currently accessible neighboring node already exists in the open node list or the closed node list, then the temporary cumulative actual cost is compared with the cumulative actual cost already recorded for the currently accessible neighboring node; if the temporary cumulative actual cost is less than the recorded cumulative actual cost, then the cumulative actual cost of the currently accessible neighboring node is updated to the temporary cumulative actual cost, and its parent node pointer is pointed to the currently to-be-expanded node; if the currently accessible neighboring node was originally in the closed node list, then it is moved back to the open node list.

[0106] A1054. After completing the expansion process for all accessible adjacent nodes of the current node to be expanded, the current node to be expanded is moved from the open node list to the closed node list. The expansion process in this embodiment refers to: for the current node to be expanded selected from the open node list, traversing each of its accessible adjacent nodes, and performing a complete sequence of operations including determining the state of the adjacent node, calculating its temporary cumulative actual cost, and deciding whether to create a new node record or update an existing node record based on the cost comparison result, including possibly moving the node from the closed node list back to the open node list.

[0107] A1055. Repeat the steps of selecting the current node to be expanded from the open node list, calculating the temporary cumulative actual cost of the passable adjacent nodes and the improved heuristic function value, comparing and updating the cost, and transitioning the node state until any of the following termination conditions are met: the node corresponding to the endpoint coordinate of the local path planning of the track is added to the closed node list, or the open node list is empty.

[0108] A106. When the node corresponding to the endpoint coordinate of the local path planning is added to the closed node list, starting from the endpoint node coordinate of the local path planning, trace back to the initial path node along its parent node pointer in sequence to generate a locally optimal track path trajectory from the starting point coordinate of the local path planning to the endpoint coordinate of the local path planning.

[0109] For example, given the coordinates of the endpoint of the local path planning of the track, the corresponding node is found and added to the list of closed nodes, denoted as node Z. The parent node pointer of node Z points to its predecessor node Y.

[0110] Algorithm startup path backtracking process:

[0111] Starting from the endpoint node Z, access its parent node pointer to obtain node Y.

[0112] Access the parent node pointer of node Y to obtain node X.

[0113] Visit the parent node pointer of each node in turn until the initial path node whose parent node pointer is null or points to itself is reached. This node is denoted as node A and corresponds to the starting coordinates of the local path planning of the track.

[0114] The node sequence A>...>X>Y>Z formed by reverse tracing is the ordered path node sequence of the locally optimal trajectory from the starting point A to the ending point Z. This sequence can be directly converted into control commands to guide the mobile robot to move along this trajectory.

[0115] A107. Collect and store the historical operation data set of the mobile robot in the track scenario and the operation data set of the initial path node in the flat ground scenario; the historical operation data includes at least the actual cargo weight corresponding to each path planning task, the generated ordered path node sequence, the actual travel time, the actual energy consumption value, and the actual safety event data recorded by the sensors.

[0116] A108. Based on the historical operation data set in the track scenario and the operation data set of the initial path nodes in the flat terrain scenario, a comprehensive performance score is calculated for each historical operation task. The comprehensive performance score is calculated by weighted summation of the following sub-indicators: path length index, task completion time index, energy consumption index, and safety violation penalty index. In this embodiment, the path length index is calculated by summing the geometric distances between all adjacent nodes in the generated ordered path node sequence; the task completion time index is calculated based on the actual travel time; the energy consumption index is calculated based on the actual energy consumption value combined with the actual cargo weight; and the safety violation penalty index is calculated based on the number and severity level of the actual safety incident data.

[0117] A109. Based on the comprehensive performance score and the historical values ​​of the parameters to be optimized in the orbital navigation strategy sub-model, a parameter loss function is constructed; the parameters to be optimized include at least: the mapping coefficient used when calculating the edge weight penalty value, the fusion coefficient used when incorporating the dynamic reserved weight value into the temporary cumulative actual cost calculation, and the weight coefficient of the potential field heuristic term when calculating the improved heuristic function value.

[0118] A110. Using a gradient descent-based optimization algorithm or a policy evaluation-based reinforcement learning algorithm, with the goal of minimizing the parameter loss function, the parameters to be optimized are adjusted in multiple iterations. In each iteration, the historical path planning task is re-simulated based on the updated parameter values, a new comprehensive performance score is calculated, and the parameters are adjusted according to the direction and magnitude of the score change, until the parameter loss function converges or the preset number of iterations is reached.

[0119] A111. The stable parameter values ​​obtained after optimization training are updated and deployed to the corresponding configuration of the orbital navigation strategy sub-model for subsequent real-time path planning tasks, so that the generated orbital path control commands tend to be globally optimal in terms of path length, time cost, energy consumption and safety.

[0120] The intelligent agriculture four-way vehicle path planning technology solution proposed in this application has significant and multi-dimensional beneficial effects. First, by accurately obtaining the weight of the loaded goods and calculating the first inertial force influence coefficient, it integrates this into the calculation of the cost of turning, lifting, and other actions. Simultaneously, by combining the modified artificial potential field method to perceive geometric constraints and obstacles such as track curves in real time, the node repulsion potential field strength is dynamically converted into edge weight penalty values ​​and superimposed on the path cost. This can proactively guide the algorithm to abandon sharp turns that hug the inside of curves and choose smoother trajectories with larger turning radii. This not only effectively reduces the interference of inertial forces on vehicle stability under full load conditions but also reduces the number of turns, significantly improving the smoothness of the track path and the safety of transporting fully loaded goods. Second, by generating a formatted query command containing scene type, node coordinates, and expected arrival timestamp, a dynamic reservation weight value is obtained from the spatiotemporal reservation table of the central scheduler. This weight value is calculated based on the aggregation value of the multi-vehicle repulsion potential field energy to estimate the conflict probability and combined with the remaining safe passage time of the node, which can accurately quantify the passage of path nodes. The system addresses the risks of collisions and time urgency, thus proactively avoiding path intersection conflicts in multi-vehicle collaborative operations within multi-layered, multi-lane greenhouse scenarios and ensuring traffic order in complex environments. Thirdly, it optimizes node search directions by constructing an improved heuristic function with potential field heuristics, combined with historical operational data including cargo weight, path sequence, and travel time to calculate a comprehensive performance score. This leads to the construction of a parameter loss function and the iterative optimization of key parameters such as edge weight mapping coefficients and dynamic weight fusion coefficients using algorithms like gradient descent. This achieves a globally optimal balance in terms of path length, task completion time, and energy consumption, significantly improving operational efficiency and reducing energy costs. Fourthly, it adaptively adjusts the planning logic based on the scene differences between the three-dimensional coordinates of the track and the two-dimensional grid coordinates of the flat ground, adapting to the mixed navigation needs of track and flat ground. Furthermore, the parameter-optimized model better adapts to the complex environment and dynamic task changes of smart agricultural greenhouses, providing high-precision and high-reliability technical support for automated material transfer and effectively contributing to the efficient and intelligent upgrading of agricultural production.

[0121] It should be further explained that the flat terrain navigation strategy sub-model in this embodiment is used to calculate the locally optimal flat terrain path trajectory in a flat terrain scenario, including:

[0122] B101. Acquire 2D grid map data of the flat terrain scene, coordinates of the starting point and ending point of the local path planning, and obstacle information. It should be further noted that the 2D grid map data of the flat terrain scene in this embodiment is collected and constructed using LiDAR or visual SLAM technology to characterize the passable areas and static structures of the environment. The coordinates of the starting point and ending point of the local path planning are distributed in real time by the upper-level task scheduler after global task decomposition, used to define the search range of the current flat terrain sub-task. Obstacle information is obtained through real-time sensors, such as LiDAR, depth cameras, or static annotations in pre-stored maps, used to calculate the repulsive field in the artificial potential field method to identify risk areas. These data collectively serve as input to the flat terrain navigation strategy sub-model, supporting the ant colony optimization algorithm combined with the artificial potential field method to search for and plan locally optimal paths on a 2D plane.

[0123] B102. Calculate a gravitational potential field value based on the distance from the grid node to the target point and a repulsive potential field value based on the distance from the grid node to the nearest obstacle for each grid node in the map.

[0124] B103. Initialize the core parameters of the ant colony optimization algorithm, including the preset ant population size, maximum number of iterations, basic pheromone volatility coefficient, pheromone enhancement constant, first heuristic factor characterizing pheromone importance, and second heuristic factor characterizing distance heuristic importance.

[0125] B104. Initialize the pheromone concentration matrix. Based on the starting coordinates and ending coordinates of the local path planning on flat terrain, and the gravitational potential field value, assign an initial pheromone concentration to each node. The initial pheromone concentration is positively correlated with the gravitational potential field value of the corresponding grid node and negatively correlated with the ending information of the previous adjacent locally optimal trajectory. Specifically:

[0126] ,in Indicates from path node i To path node j The initial pheromone concentration along the path; Let Q be the initial baseline pheromone, and Q be the pheromone enhancement coefficient. This represents the gravitational potential field adjustment coefficient, which is a multiplier coefficient used to adjust path nodes. j The gravitational potential energy value affects the initial pheromone concentration The intensity of the impact; The larger the value, the stronger the attenuation effect of the gravitational potential energy on the initial pheromone; This represents the historical best path length or reference path cost passed from the best trajectory path in the previous adjacent orbit scenario. Represents path nodesj The gravitational potential energy value, calculated using the artificial potential field (APF) method, reflects the degree of attraction of the node towards the target. This represents an exponential function with the natural constant as its base. The design intent of the initial pheromone concentration matrix in this embodiment is to assign differentiated initial pheromones to each path before the ant colony algorithm begins, based on global path information and potential field inspiration. This enhances the pheromone concentration along historically superior paths and around nodes with high gravitational potential energy, thereby guiding the search process to converge more quickly to a safe and efficient feasible path.

[0127] B105. In each iteration, a path search is performed independently for each virtual ant; the path search is conducted in a network composed of the two-dimensional raster map data of the flat terrain scene, where each raster node is regarded as a path node, and the path search is specifically performed as follows:

[0128] Based on the ant's current path node, calculate the state transition probability of it moving to the next candidate path node. The state transition probability includes the pheromone concentration corresponding to the candidate path node, a heuristic value based on the distance between path nodes, and the repulsive potential field value of the candidate path node. The repulsive potential field value serves as a penalty factor to reduce the probability of a grid node near an obstacle being selected. The state transition probability is: ,in This represents the probability that the k-th ant moves from path node i to path node j at time t. This represents the pheromone concentration along the path from node i to node j at time t. This represents the heuristic information from path node i to path node j at time t. ,in Let be the repulsive potential energy value at path node j calculated by the Artificial Potential Field (APF) method. The repulsion sensitivity coefficient is set by those skilled in the art based on historical experience; α As the first heuristic factor characterizing the importance of pheromones, β As a second heuristic factor characterizing the importance of distance heuristics; This indicates that at time t, the path from path node i to path node... l The inspirational information above This indicates that at time t, the path from path node i to path node... l The concentration of pheromones on the surface This is the set of all possible next nodes that the k-th ant can choose when it is located at path node i at time t. This set defines the feasible movement range of the ant and is usually determined by the algorithm rules and environmental constraints. It mainly includes: unvisited nodes: according to the tabu list mechanism of the ant colony algorithm, Nodes already visited by ant k on the current path are excluded to avoid forming a cycle; walkable nodes: in the path planning problem, it only includes adjacent nodes that can be directly and safely reached from node i on the current path. It should be further explained that the core principle of constructing this state transition probability in this embodiment lies in calculating the repulsive potential energy value using the Artificial Potential Field (APF) method. Heuristic information embedded as a penalty item This paper improves the probabilistic selection mechanism of traditional ant colony optimization algorithms. The motivation is to overcome the shortcomings of traditional methods in real-time obstacle avoidance in dynamic environments, enabling ants to not only rely on "collective experience" accumulated through pheromones during path exploration, but also to perceive and avoid obstacle-threatening areas in real time. The ultimate effect is that the algorithm achieves a balance between global optimization and local obstacle avoidance, guiding ants to choose nodes closer to obstacles with a lower probability, thereby planning safer, smoother, and collision-free paths that adapt to dynamic environments in complex flat terrain scenarios.

[0129] Based on the calculated state transition probabilities of each candidate path node, the roulette wheel selection method is used to determine the next moving path node of the ant, and the selected path node is added to the current ant's current path sequence.

[0130] Repeat the state transition and path node selection process until the ant reaches the local path planning endpoint coordinates on flat ground or the path construction fails.

[0131] B106. Repeat the iterative process consisting of the ant path iterative construction step and the pheromone global update and collaborative adjustment step until the preset maximum number of iterations is reached or the pheromone distribution converges. Specifically:

[0132] The steps for iteratively constructing an ant path include: independently performing a complete path search from the starting point to the ending point for each virtual ant in the ant colony;

[0133] The steps for global pheromone update and collaborative adjustment include: updating the pheromone concentration of the entire graph according to predefined rules based on the paths constructed by all ants in the current iteration and the dynamic reservation weight values ​​obtained from the spatiotemporal reservation table;

[0134] In this embodiment, the global pheromone update uses the following formula:

[0135] ,in This represents the updated pheromone concentration value on the path edge (i,j) in this iteration. It is the result state after the iteration, not the increment itself. It is the key to the algorithm's group learning and is used to strengthen high-quality paths. The total length or total cost of the path constructed by the k-th ant in this iteration is the core indicator for evaluating the quality of a single ant's path. The smaller the value, the better the path found by the ant, and therefore the more pheromone rewards it will receive along the path edges. This represents the spatiotemporal reservation weight, which is a dynamic value returned by the spatiotemporal reservation table based on the query. It quantifies the risk of passage conflict and time urgency of node j at the expected arrival time t. The larger the value, the higher the risk of the node being occupied at the future time t or the tighter the available time window. The maximum reserved weight is a preset normalization constant used to allocate dynamic spatiotemporal reserved weights. Scaling to a reasonable range, typically compared to 1; The pheromone evaporation coefficient at time t controls the proportion of the original pheromone on the path edge (i,j) that is retained after the iterative update. The smaller the size, the more pheromones are retained, and the longer the influence of historical experience lasts; conversely, The larger the size, the higher the pheromone retention rate. The smaller the value, the weaker the influence of historical information, and the stronger the algorithm's ability to explore new paths. The base evaporation coefficient is the benchmark value for pheromone evaporation, which determines the standard pheromone evaporation rate under conditions without additional risks of spatiotemporal conflict. The adjustment factor is a weighting coefficient used to control the degree to which the spatiotemporal reservation weights adjust the pheromone evaporation coefficient. The larger the value, The more significant the effect on the evaporation rate, the more pronounced the effect. This represents the total amount of pheromone added by all ants on the path edge (i,j) in this iteration; this embodiment introduces... This dynamic parameter from the spatiotemporal collaborative scheduling layer has profoundly modified the core update mechanism of the traditional ACO; in the pheromone increment stage, the formula... Even if the geometric length of a path The path is very short; if the nodes it passes through have a high risk of conflict at a specific time, the pheromone enhancement gained through this path will be significantly suppressed; pheromone evaporation coefficient Weights will be reserved based on time and space. Dynamic adjustment, its core meaning lies in using formulas This translates the high conflict risk of nodes into an increase in pheromone evaporation rate; specifically, it involves the spatiotemporal reservation weight of a node at a specific moment. When the value is higher, the formula will calculate a larger value. The intention is to make the old pheromones on the path associated with the node evaporate faster and the attraction weaken earlier, thereby guiding subsequent ants to actively avoid these high-risk spatiotemporal nodes. This design effectively avoids the premature convergence problem caused by excessively high pheromone concentration in the traditional ACO algorithm.

[0136] To more clearly illustrate the process described in B101-B106 above, the following example is provided:

[0137] After Robot_A switches from the track to flat ground, its flat ground navigation strategy sub-model quickly generates an initial optimized path from E1 to G1. During travel, when the system suddenly detects a temporarily placed container ahead (corresponding to a dynamic obstacle) via sensors, the APF repulsive field is updated in real time. The state transition probability of the nodes in this area immediately decreases in the ant's next path selection, guiding Robot_A to adjust its local path in real time to bypass the obstacle.

[0138] Meanwhile, the central scheduler's spatiotemporal reservation table predicts that another empty Robot_B will pass through the intersection node N1 in 15 seconds. This information is converted into a higher reservation weight value and fed back to Robot_A's ant colony algorithm. In the subsequent pheromone update, although the path through node N1 may be shorter, the pheromone increment it receives is significantly suppressed, and the old pheromone on node N1 evaporates more rapidly. Ultimately, Robot_A's ant colony algorithm converges to a slightly longer but suboptimal path that completely avoids node N1, thus achieving safe and efficient collaboration with Robot_B through time-shifted and spatially staggered operations, improving the overall system throughput.

[0139] This example clearly demonstrates that by embedding the real-time obstacle avoidance capability of the APF (Advanced Persistent Threat) into every decision made by the ants, and by integrating dynamic weights reflecting the future spatiotemporal conflict risk of multiple vehicles into the ant colony's collective learning mechanism, the flat-ground navigation strategy sub-model of this application can simultaneously achieve the two objectives of real-time obstacle avoidance and multi-vehicle collaboration in complex dynamic flat-ground scenarios. Furthermore, the fusion of historical track information and gravitational potential field during the initialization phase ensures the smoothness and continuity of the transition from track-based to flat-ground navigation strategy.

[0140] B107. From all the paths generated in all iterations, select the path with the highest comprehensive evaluation as the optimal solution for this round of planning; whereby the comprehensive evaluation of the path is determined based on the path length, smoothness index, and safety index.

[0141] It should be further explained that in this embodiment, the path with the highest overall evaluation is selected as the optimal solution for this round of planning, including:

[0142] B1071. Based on the complete candidate path independently constructed by each virtual ant in this iteration, obtain the path node sequence set of all candidate paths; the path node sequence set records all ordered grid nodes traversed by the ant from the starting coordinates of the local path planning on flat ground to the ending coordinates of the local path planning on flat ground.

[0143] B1072. For each candidate path in the path node sequence set, perform multi-dimensional cost calculation sequentially; the multi-dimensional cost calculation includes:

[0144] First-dimensional cost calculation: Based on the path node sequence of the candidate path, the Euclidean distance between all adjacent nodes is accumulated to obtain the total geometric length cost of the candidate path;

[0145] Second dimension cost calculation: Traverse the path node sequence of the candidate path, identify all turning points in the path where the driving direction changes, and calculate the total turning penalty cost of the path based on the preset unit turning penalty coefficient.

[0146] Third-dimensional cost calculation: Based on the pre-generated artificial potential field map, query the repulsive potential field value corresponding to each node in the path node sequence of the candidate path, and accumulate them to obtain the total potential field repulsive cumulative cost of the path.

[0147] Fourth dimension cost calculation: Based on the pre-set spatiotemporal reservation table, the path node sequence of the candidate path and its corresponding expected arrival time window are used as query input to obtain the continuous value reservation weight corresponding to each node, and the reservation weights of all nodes on the path are weighted averaged or accumulated to obtain the total spatiotemporal conflict risk cost of the path.

[0148] B1073. Based on preset weighting coefficients, the total geometric length cost, total turning penalty cost, total potential field repulsion accumulation cost, and total spatiotemporal conflict risk cost of each candidate path are weighted and fused to calculate the comprehensive evaluation score of the candidate path; wherein, the comprehensive evaluation score is negatively correlated with the sum of the costs, that is, the higher the cost, the lower the score.

[0149] B1074. Compare the comprehensive evaluation scores of all candidate paths in this iteration, and use the maximum value retrieval algorithm to select the only candidate path with the highest comprehensive evaluation score from the set of path node sequences.

[0150] B1075. The candidate path with the highest comprehensive evaluation score is determined and output as the optimal solution path for this round of ant colony optimization iteration.

[0151] B108. The optimal decoding is converted into an ordered sequence of path points consisting of a series of continuous two-dimensional coordinate points, which is output as the local optimal flat terrain path trajectory.

[0152] B109. Collect and store the historical operational data set of the mobile robot in flat terrain scenarios, as well as the endpoint path information of the locally optimal trajectory in the adjacent previous track scenario. The data includes at least the actual path sequence, completion time, energy consumption, and safety event records for each task. It should be further noted that the endpoint path information of the locally optimal trajectory in the adjacent previous track scenario in this embodiment is primarily used as the initial state input and continuity constraint for path planning in flat terrain scenarios. Specifically, this endpoint path information includes the precise coordinates of the endpoint, the direction of motion upon arrival, and the velocity vector. This information is provided to the flat terrain navigation strategy sub-model to determine the starting point of the local path planning in flat terrain and to constrain the initial search direction, thereby ensuring a smooth and abrupt transition of the path from the end of the track scenario to the starting point of the flat terrain scenario in terms of geometry and motion state. This avoids drastic changes in path direction or re-acceleration / deceleration caused by scene switching, ensuring the overall smoothness and optimal energy consumption of cross-scenario navigation.

[0153] B110. Based on the historical operational data set, a comprehensive performance score is calculated for each historical task. This score is obtained by weighted summation of the path length score, time efficiency score, energy consumption score, and safety score. In this embodiment, the path length score is obtained by accumulating and evaluating the geometric distances between all adjacent nodes in the ordered path point sequence; the time efficiency score is obtained by evaluating the actual travel time of the mobile robot in completing the task; the energy consumption score is calculated by combining the actual energy consumption value of the mobile robot with the actual cargo weight during task execution; and the safety score is calculated by the number of events recorded in the safety event log and the preset severity level corresponding to each event. The four-dimensional scores are finally fused through preset weighting coefficients to form a single comprehensive performance score.

[0154] B111. Construct a loss function with the comprehensive performance score as the optimization objective. The independent variables of the loss function include key adjustable parameters from the ant colony optimization algorithm. These key adjustable parameters include at least a first heuristic factor, a second heuristic factor, a pheromone evaporation coefficient, a pheromone enhancement constant, and a repulsion sensitivity coefficient of the repulsion potential field penalty term in the state transition probability. γ ;

[0155] B112. Using a gradient descent optimization algorithm or a policy gradient-based reinforcement learning algorithm, with the goal of minimizing the loss function, the key adjustable parameters are iteratively adjusted in multiple rounds. In each round of adjustment, the updated parameters are used to simulate historical tasks and calculate a new comprehensive performance score, and the parameters are updated accordingly.

[0156] B113. Update and deploy the stable parameter set obtained after training convergence to the flat ground navigation strategy sub-model for subsequent real-time path planning tasks.

[0157] This embodiment integrates ant colony optimization algorithm and artificial potential field method, enabling the algorithm to simultaneously possess the global exploration capability of swarm intelligence and the local fine-grained obstacle avoidance capability of artificial potential field for real-time obstacles. This significantly improves the robustness and path safety in dealing with dynamic obstacles in unstructured flat terrain scenarios. Furthermore, by introducing a spatiotemporal reservation table maintained by a central scheduler and designing dynamic reservation weight values, the risk of space occupation conflict and the urgency of time windows at future moments are quantified into calculable path costs. This allows each virtual ant to proactively avoid high-risk nodes during path exploration, achieving proactive collaboration and conflict resolution among multiple mobile robots from the root of swarm intelligence, greatly improving the overall system throughput and operational safety. Crucially, the flat terrain navigation strategy sub-model is initialized... The pheromone concentration matrix incorporates the endpoint information of the optimal path in the previous orbital scenario as a constraint. Under the scheduling of the switching prediction sub-model, pre-transformation and pre-loading of potential field data and model parameters ensure a smooth and seamless transition in both spatial geometry and motion state between the navigation strategy and the terrain, eliminating path abrupt changes and efficiency losses caused by scene switching. Finally, by continuously collecting historical operational data and constructing a loss function with a comprehensive performance score as the objective, gradient descent and other optimization algorithms are used to iteratively learn and self-adjust the key adjustable parameters of the ant colony optimization algorithm. This enables the entire navigation strategy to possess online adaptive optimization capabilities, continuously approaching the global optimal performance under different loads and environmental complexities, achieving a continuous and synergistic improvement in system efficiency, safety, and energy consumption. Therefore, through this series of interconnected and progressively advancing technical means, this invention ultimately achieves a systematic breakthrough in the adaptiveness, coordination, smoothness, and overall optimality of mobile robot path planning in mixed and complex environments.

[0158] It should be further explained that the construction process of the switching prediction sub-model in this embodiment includes:

[0159] C101. Based on the real-time position coordinates, real-time running speed of the mobile robot, and the transition connection point coordinates included in the scene transition point parameters, calculate the estimated remaining time required to reach the transition connection point coordinates from the real-time position coordinates. In this embodiment, to ensure the continuity of navigation control commands when the mobile robot crosses the boundary of a mixed track and flat ground scene, it is necessary to accurately determine the optimal timing for algorithm switching in advance to avoid running pauses, trajectory discontinuities, and control instability caused by switching delays or premature switching. The principle is to construct a forward-looking time estimator based on real-time motion state: First, continuously receive the real-time position coordinates and real-time running speed vector from the mobile robot positioning system. Secondly, the coordinates of the transition point related to the current task are retrieved from the pre-stored map data. Then, by calculating the real-time Euclidean distance between the real-time position coordinates and the coordinates of the transition point, and by performing decoupling analysis in conjunction with the magnitude and direction of the current running speed vector, the optimal estimated remaining time required to reach the transition point in the current motion state is dynamically calculated. This time value serves as a dynamic and precise triggering benchmark, directly driving a series of preparation processes such as cross-scene data pre-conversion and model parameter pre-loading. Thus, before physically reaching the scene boundary, a smooth transition of the navigation system is completed at the logical and data levels, ultimately ensuring the consistency of the global path trajectory and the seamless connection of control commands at the temporal level.

[0160] C102. Based on a preset switching delay and safety margin, a switching advance threshold is set; when the estimated remaining time is equal to or less than the switching advance threshold, the current moment is determined to be the optimal trigger time for navigation strategy switching; the steps for constructing the switching advance threshold in this embodiment include: first, through offline testing and online statistics, obtaining the average time and maximum fluctuation duration required for the navigation strategy sub-model to complete data pre-conversion, parameter initialization, and seamless handover of control commands under typical load and standard scenarios, and using the maximum fluctuation duration as the inherent switching delay benchmark value of the system; second, dynamically calculating the adaptive safety margin value, which is based on... The system considers three real-time variables: the real-time weight of the goods carried by the mobile robot obtained through a weighing sensor and converted into the corresponding first inertial force influence coefficient; the scene complexity evaluation value calculated in real-time by analyzing the obstacle density and path curvature in the current grid map; and the average safety event interval time under recent similar scenarios obtained by querying the historical operation database. Finally, a linear weighted fusion method is used to add the switching delay benchmark value to the adaptive safety margin value calculated by the first inertial force influence coefficient, the scene complexity evaluation value, and the average safety event interval time through a preset formula. The sum is then used to construct the switching lead threshold.

[0161] C103. Before reaching the optimal trigger time point, initiate the cross-scenario artificial potential field data conversion process, specifically as follows:

[0162] When the predicted next target scene is a flat terrain scene, a two-dimensional artificial potential field data pre-generation operation is performed, including: acquiring the three-dimensional artificial potential field data constructed under the current orbit scene; using a preset oblique projection formula, mapping the three-dimensional repulsive field values ​​in the three-dimensional artificial potential field data to the two-dimensional plane corresponding to the flat terrain scene; and simultaneously generating an initial repulsive field dataset for the flat terrain scene suitable for two-dimensional raster maps by assigning corresponding projection weights to the three-dimensional repulsive field values ​​at different height layers and performing vector synthesis. The oblique projection formula is as follows: ,in This represents the two-dimensional planar coordinates in the two-dimensional flat scene grid map obtained after projection calculation. This is the target output for cross-scene potential field data conversion. These represent the three-dimensional spatial coordinates of a path node or obstacle repulsive field source point in a three-dimensional orbit scene. x and y represent the position in the horizontal plane, z represents the vertical height layer, such as Z1, Z2, and Z3 orbit layers. θ is the projection angle, a crucial projection parameter, typically set to a fixed value such as 45° or 30°. It determines the "tilt" at which the height information in three-dimensional space is projected onto the two-dimensional plane, affecting the shape and scale of the two-dimensional potential field distribution after projection. and These are the projection direction angles along the X and Y axes, respectively. They control which horizontal direction the 3D spatial coordinates are tilted during projection, ensuring the projected 2D map aligns with the actual orientation of the flat terrain scene. The tilted projection formula in this embodiment is the core technical means for achieving cross-scene artificial potential field data pre-conversion in this application by switching the prediction sub-model. Its purpose is to pre-convert and synthesize the calculated 3D obstacle repulsion field data from the track map into initial 2D potential field values ​​usable for flat terrain navigation, in a way that preserves spatial relative relationships and threat levels, before the mobile robot switches from a 3D track scene to a 2D flat terrain scene. This is achieved by introducing the projection angle θ and the direction angle. and This formula simulates the axonometric projection principle in engineering drawings, enabling obstacle information at different height levels to be reasonably mapped onto the plane, thereby ensuring the continuity and consistency of navigation environment perception information during scene switching; wherein, the superposition of the three-dimensional repulsive force value on the two-dimensional plane in this embodiment is: ,in This indicates that after projection and overlay, the coordinates in the 2D flat ground scene raster map are... The synthesized two-dimensional repulsive potential field value obtained at the grid node; this is the output of the formula, which will be directly used as the input of the Artificial Potential Field (APF) method in the flat ground navigation strategy sub-model to identify the obstacle threat intensity at that location; z∈H indicates that the index variable z traverses all valid track height layers, and H represents the number of track height layers. In the application context of multi-layer track systems in smart agricultural greenhouses, tracks are planned to run at different vertical heights, and these height layers are systematically numbered, such as Z1, Z2, and Z3 layers. The set H formally defines all height layers that need to be included in the calculation scope of the system. For example, if the system contains three track layers, then H={1,2,3}, corresponding to Z1, Z2, and Z3 layers in physical space, respectively; for example, Z1, Z2, and Z3 layers in a smart agricultural greenhouse. This limits the scope of the summation, ensuring that only spatial layers with actual obstacles or tracks are included; This represents the original three-dimensional repulsive potential energy value located at three-dimensional coordinates (x,y,z) in a three-dimensional track scene. This value is calculated by the APF algorithm in the track scene based on the distance to obstacles, such as track curves, other vehicles, and fixed facilities. This represents the weighting coefficient for the height layer, a weight value associated with a specific height layer z. Its function is to adjust the contribution ratio of the three-dimensional repulsive potential energy values ​​of different height layers to the final two-dimensional plane synthesized potential field. For example, the track layer closer to the target flat ground plane (such as layer Z1) may be assigned a higher weight because it has a more direct impact on potential spatial conflicts when traveling on flat ground. In this embodiment, under the specific application background of multi-layer track and flat ground operation in a smart agricultural greenhouse, this formula is a key technical link to achieve continuous environmental perception and ensure navigation safety without blind spots. Its refined process is as follows: when the system predicts that the four-way vehicle will drive from the three-dimensional track into the two-dimensional flat ground area, the prediction sub-model is switched. First, the coordinates of the three-dimensional repulsive field source points of obstacles on the track at each height layer, such as other running vehicles and track support structures, are mapped to the two-dimensional plane of flat ground through the oblique projection formula. Then, this oblique projection formula is activated as the core synthesizer to fuse multiple three-dimensional repulsive potential energy values ​​from different height layers that fall on the same two-dimensional grid node after projection. The system dynamically assigns differentiated weight coefficients to each height layer based on its spatial vertical distance from the target flat ground plane, obstacle type, and historical conflict probability. For example, Z1 layer track obstacles, which have a greater intersection with the flat ground space and higher vehicle density, are given higher weights, while the weights of higher Z3 layer obstacles are correspondingly reduced. Finally, through weighted superposition calculations, a two-dimensional initial repulsive field is generated that comprehensively reflects the combined threat of multiple layers of obstacles overhead. This allows the flat ground navigation algorithm to proactively avoid potential intersection conflicts from the overhead track area from the very beginning of its path search. Thus, in complex grade-separated operation scenarios, it achieves lossless inheritance and intelligent fusion of environmental threat information during dimensional transformations, fundamentally avoiding perception gaps and safety risks caused by scene switching.

[0163] C104. The coordination and initialization of navigation model parameters are performed in parallel with the cross-scene artificial potential field data conversion step, specifically including:

[0164] When the predicted next target scene is a flat scene, the first path geometric information from the current position of the mobile robot to the coordinates of the transition connection point is received from the current track navigation strategy sub-model.

[0165] Based on the first path geometry information, the pheromone concentration matrix of the ant colony optimization algorithm in the flat ground navigation strategy sub-model is initialized in a biased manner, so that the grid nodes located in the direction indicated by the first path geometry information obtain the updated initial pheromone concentration.

[0166] When the predicted next target scenario is a track scenario, the path optimization feedback data output by the current flat terrain navigation strategy sub-model is received; the path optimization feedback data includes at least the key node sequence or curvature distribution information that characterizes the geometric smoothness of the path, generated during the iterative solution of the flat terrain scenario.

[0167] Based on the path optimization feedback data, adjust the improved A in the orbital navigation strategy sub-model. The improved algorithm's heuristic function relies on the gravitational potential field adjustment coefficient corresponding to the potential field heuristic term, so as to guide the generated local optimal trajectory path to be consistent and continuous with the smoothness represented by the path optimization feedback data at the connection point.

[0168] C105. When the mobile robot actually runs to the coordinates of the transition point, the central scheduler switches the currently active navigation strategy sub-model to the target navigation strategy sub-model that has completed data and parameter preloading according to the navigation strategy identifier. The target navigation strategy sub-model uses the pre-transformed initial potential field dataset and pre-loaded optimization parameters to generate continuous control instructions for the new scenario, thereby achieving seamless switching of navigation strategies.

[0169] The precise, forward-looking, and collaborative switching mechanism constructed by the switching prediction sub-model described in this embodiment solves the problems of pauses, trajectory abrupt changes, and perception discontinuities that exist in navigation algorithms when switching at the boundaries of mixed scenes. Its effectiveness stems first from temporal accuracy. By dynamically calculating and estimating the remaining time through the fusion of real-time pose and map information, and combining this with a transition lead threshold adaptively based on load and scene complexity for precise triggering, the system ensures that it completes all preparations just before physically reaching the boundary. This eliminates operational interruptions and speed losses caused by algorithm switching delays or premature switching, guaranteeing the continuity and time optimization of mobile robot operations. Secondly, it is reflected in the continuity of environmental perception. By using an oblique projection formula to map the three-dimensional orbital repulsive field onto a two-dimensional plane, and weighting and fusing it according to the height layer and spatial threat level, a two-dimensional initial potential field containing multi-layered spatial obstacle information is generated. This allows the flat-ground navigation algorithm to perceive and avoid potential conflicts from the overhead orbital area immediately upon startup, achieving lossless inheritance and fusion of safety information during dimensional transformation, eliminating perception blind spots and safety risks caused by scene switching. This further enhances the smoothness of strategy coordination. By performing parallel initialization of model parameters before switching, such as biased initialization of ant colony pheromones on flat ground using track path geometry information, or optimizing the track heuristic function based on the smoothness feedback of flat ground paths, the path search between the preceding and following scenes becomes naturally coherent in terms of geometric direction and motion trend at the connection point. This avoids drastic changes in path direction and unnecessary re-acceleration and deceleration, achieving a smooth transition at the control command level. Ultimately, these technical effects converge into significant system-level gains in complex dynamic scenarios such as smart agricultural greenhouses where multiple tracks and flat ground intertwine. This enables multiple mobile robots to efficiently perform their respective transport tasks while achieving safe, orderly, and smooth collaborative operation and scene traversal in both spatial and temporal dimensions, significantly improving the overall system's operational efficiency, reliability, and intelligence level.

[0170] Example 2:

[0171] Please see Figure 2 Another embodiment of the present invention provides a mobile robot path optimization method supporting track mode switching, comprising:

[0172] S1. The mobile robot's operating scenario and real-time load status are comprehensively encoded to generate navigation task seeds;

[0173] S2. Based on the navigation task seed, select the corresponding navigation strategy model and combine it with the preset spatiotemporal reservation table to generate control instructions for the executable trajectory, and output the instruction readiness indication to the central scheduler; the navigation strategy model includes a track navigation strategy sub-model, a flat ground navigation strategy sub-model, and a switching prediction sub-model; the spatiotemporal reservation table is used to calculate a continuous value reservation weight associated with the probability of multi-vehicle conflict and the remaining safe passage time of the path node for each path node based on the three-dimensional coordinates or two-dimensional grid coordinates of each path node in the executable trajectory and its corresponding time window, so as to perform multi-vehicle collaborative conflict avoidance and resource scheduling for the selected track navigation strategy sub-model or flat ground navigation strategy sub-model;

[0174] S3. After receiving the instruction ready indication, the control instruction for the executable trajectory is sent to the corresponding track or flat ground navigation control loop.

[0175] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the claims. All of these variations are within the protection scope of the present invention.

Claims

1. A mobile robot path optimization system supporting track mode switching, characterized in that, include: The encoding module comprehensively encodes the mobile robot's operating scenario and real-time load status to generate navigation task seeds; The instruction generation module, based on the navigation task seed, selects the corresponding navigation strategy model and combines it with a preset spatiotemporal reservation table to generate control instructions for the executable trajectory, and outputs an instruction readiness indication to the central scheduler. The navigation strategy model includes a track navigation strategy sub-model, a flat ground navigation strategy sub-model, and a switching prediction sub-model. The spatiotemporal reservation table is used to calculate a continuous value reservation weight for each path node based on the three-dimensional coordinates or two-dimensional grid coordinates of each path node in the executable trajectory and its corresponding time window. This weight is associated with the probability of multi-vehicle conflict and the remaining safe passage time of the path node, so as to perform multi-vehicle collaborative conflict avoidance and resource scheduling for the selected track navigation strategy sub-model or flat ground navigation strategy sub-model. Upon receiving the instruction ready indication, the control module sends the control instruction for the executable trajectory to the corresponding track or flat ground navigation control loop.

2. The mobile robot path optimization system supporting track mode switching as described in claim 1, characterized in that, The orbital navigation strategy sub-model is derived from improved A. The algorithm is constructed by combining an improved APF and is used to calculate the local optimal trajectory in a structured track scenario. The flat ground navigation strategy sub-model is constructed by combining the ACO algorithm with APF and is used to calculate the local optimal flat ground path trajectory in a flat ground scenario. The switching prediction sub-model is used to calculate the optimal time point for navigation strategy switching and the remaining time length for navigation strategy switching based on the real-time position coordinates of the mobile robot in the corresponding scenario and the scene switching point parameters in the pre-stored map data. The navigation task seed specifically includes: a path cost vector, which at least includes the actual distance from the starting point to the current estimated node, the estimated number of turns, and a first inertial force influence coefficient calculated based on the real-time load state of the mobile robot; a navigation strategy identifier, used to indicate the currently selected navigation strategy model, whose value range includes at least a first identifier value for the track scene and a second identifier value for the flat ground scene; and a scene transition point parameter, which includes the three-dimensional coordinate position at the boundary between the track scene and the flat ground scene obtained from the pre-stored map data, and the estimated remaining time to reach the boundary calculated in real time based on the current speed of the mobile robot and the three-dimensional coordinate position.

3. The mobile robot path optimization system supporting track mode switching as described in claim 2, characterized in that, The construction process of the orbital navigation strategy sub-model includes: Obtain the current weight of the robot's load, structured environment map data, and the coordinates of the starting point and ending point of the local path planning. Based on the current weight of the cargo loaded on the robot, and combined with the predefined full-load reference inertial force, the first inertial force influence coefficient is calculated. Initialize a configured list of open nodes and a list of closed nodes; encapsulate the starting coordinates of the local path planning of the track into an initial path node, the attributes of which include at least node coordinates, cumulative actual cost, and heuristic cost; set the cumulative actual cost of the initial path node to an initial value and add it to the list of open nodes; The node with the smallest comprehensive evaluation function value is selected from the list of open nodes as the current node to be expanded; for each traversable adjacent node of the current node to be expanded, a temporary cumulative actual cost is calculated; the temporary cumulative actual cost is constructed from at least the base distance cost, the turning action cost, the elevation action cost, the first inertial force influence coefficient, the preset edge weight penalty value, and the dynamic reservation weight value determined by querying the spatiotemporal reservation table.

4. The mobile robot path optimization system supporting track mode switching as described in claim 3, characterized in that, The construction process of the orbital navigation strategy sub-model also includes: For each of the accessible adjacent nodes, after calculating the temporary cumulative actual cost, the following logical judgment and update operations are performed: Determine whether the passable adjacent node has been recorded in the closed node list or the open node list; If the accessible adjacent node is not recorded, a node record is created for it, its cumulative actual cost is set to the temporary cumulative actual cost, its parent node pointer is pointed to the current node to be expanded, and the improved heuristic function value of the current accessible adjacent node is calculated based on the node coordinates of the current accessible adjacent node, the global target point coordinates, and the environmental potential field constructed by the sensor or map. The node record of the current accessible adjacent node is then added to the list of open nodes. If the currently accessible neighboring node already exists in the open node list or the closed node list, then the temporary cumulative actual cost is compared with the cumulative actual cost already recorded for the currently accessible neighboring node; if the temporary cumulative actual cost is less than the recorded cumulative actual cost, then the cumulative actual cost of the currently accessible neighboring node is updated to the temporary cumulative actual cost, and its parent node pointer is pointed to the currently expandable node; if the currently accessible neighboring node was originally in the closed node list, then it is moved back to the open node list.

5. The mobile robot path optimization system supporting track mode switching as described in claim 4, characterized in that, The construction process of the orbital navigation strategy sub-model also includes: After completing the expansion process for all accessible adjacent nodes of the current node to be expanded, the current node to be expanded is moved from the open node list to the closed node list; Repeat the steps of selecting the current node to be expanded from the list of open nodes, calculating the temporary cumulative actual cost and improved heuristic function value of the passable adjacent nodes, comparing and updating the cost, and transitioning the node state until any of the following termination conditions are met: the node corresponding to the endpoint coordinate of the local path planning of the track is added to the list of closed nodes, or the list of open nodes is empty; When the node corresponding to the endpoint coordinate of the local path planning is added to the list of closed nodes, starting from the endpoint node coordinate of the local path planning, the path is traced backward along its parent node pointer to the initial path node, generating a locally optimal track path trajectory from the starting coordinate of the local path planning to the endpoint coordinate of the local path planning.

6. The mobile robot path optimization system supporting track mode switching as described in claim 2, characterized in that, The flat terrain navigation strategy sub-model is used to calculate the locally optimal flat terrain path trajectory in a flat terrain scenario, including: Acquire 2D grid map data of flat terrain scene, coordinates of the starting point and ending point of local path planning on flat terrain, and obstacle information; For each grid node in the map, calculate a gravitational potential field value based on the distance from the grid node to the target point, and a repulsive potential field value based on the distance from the grid node to the nearest obstacle; The core parameters for initializing the ant colony optimization algorithm include the preset ant population size, maximum number of iterations, basic pheromone volatility coefficient, pheromone enhancement constant, first heuristic factor characterizing pheromone importance, and second heuristic factor characterizing distance heuristic importance. An initial pheromone concentration matrix is ​​initialized. Based on the starting coordinates of the local path planning on flat ground, the ending coordinates of the local path planning on flat ground, and the gravitational potential field value, an initial pheromone concentration is assigned to each node. The initial pheromone concentration is positively correlated with the gravitational potential field value of the corresponding grid node and negatively correlated with the ending information of the previous adjacent local optimal orbit path trajectory.

7. The mobile robot path optimization system supporting track mode switching as described in claim 6, characterized in that, The flat terrain navigation strategy sub-model, used to calculate the locally optimal flat terrain path trajectory in a flat terrain scenario, also includes: In each iteration, a path search is performed independently for each virtual ant; the path search is conducted in a network composed of the two-dimensional raster map data of the flat terrain scene, where each raster node is regarded as a path node, and the path search is specifically performed as follows: Based on the ant's current path node, calculate the state transition probability of it moving to the next candidate path node; the state transition probability includes the pheromone concentration corresponding to the candidate path node, the heuristic value based on the distance between path nodes, and the repulsive potential field value of the candidate path node, wherein the repulsive potential field value is used as a penalty factor to reduce the probability of grid nodes near obstacles being selected. Based on the calculated state transition probabilities of each candidate path node, the roulette wheel selection method is used to determine the next moving path node of the ant, and the selected path node is added to the current ant's current path sequence. Repeat the state transition and path node selection process until the ant reaches the local path planning endpoint coordinates on flat ground or the path construction fails.

8. The mobile robot path optimization system supporting track mode switching as described in claim 7, characterized in that, The flat terrain navigation strategy sub-model, used to calculate the locally optimal flat terrain path trajectory in a flat terrain scenario, also includes: Repeat the iterative process consisting of the ant path iterative construction step and the pheromone global update and collaborative adjustment step until the preset maximum number of iterations is reached or the pheromone distribution converges, specifically: The steps for iteratively constructing an ant path include: independently performing a complete path search from the starting point to the ending point for each virtual ant in the ant colony; The steps for global pheromone update and collaborative adjustment include: updating the pheromone concentration of the entire graph according to predefined rules based on the paths constructed by all ants in the current iteration and the dynamic reservation weight values ​​obtained from the spatiotemporal reservation table; From all the paths generated in all iterations, the path with the highest comprehensive evaluation is selected as the optimal solution for this round of planning; the comprehensive evaluation of the path is determined based on path length, smoothness index, and safety index. The optimal decoding is converted into an ordered sequence of path points consisting of a series of continuous two-dimensional coordinate points, which is output as the local optimal flat terrain path trajectory.

9. The mobile robot path optimization system supporting track mode switching as described in claim 8, characterized in that, The construction process of the switching prediction sub-model includes: Based on the real-time position coordinates, real-time running speed of the mobile robot, and the coordinates of the transition connection point included in the scene transition point parameters, the estimated remaining time from the real-time position coordinates to the transition connection point coordinates is calculated. Based on a preset switching delay and safety margin, a switching lead threshold is set; when the estimated remaining time is equal to or less than the switching lead threshold, the current moment is determined to be the optimal trigger time for navigation strategy switching. Before reaching the optimal trigger time, initiate the cross-scenario artificial potential field data conversion process, specifically as follows: When the predicted next target scene is a flat scene, a two-dimensional artificial potential field data pre-generation operation is performed, including: acquiring the three-dimensional artificial potential field data constructed under the current orbit scene, using a preset oblique projection formula to map the three-dimensional repulsive field values ​​in the three-dimensional artificial potential field data to the two-dimensional plane corresponding to the flat scene, and generating an initial repulsive field dataset for the flat scene suitable for two-dimensional raster maps by assigning corresponding projection weights to the three-dimensional repulsive field values ​​of different height layers and performing vector synthesis.

10. The mobile robot path optimization system supporting track mode switching as described in claim 9, characterized in that, The construction process of the switching prediction sub-model also includes: The coordination and initialization of navigation model parameters are performed in parallel with the cross-scene artificial potential field data conversion step, specifically including: When the predicted next target scene is a flat scene, the first path geometric information from the current position of the mobile robot to the coordinates of the transition connection point is received from the current track navigation strategy sub-model. Based on the first path geometry information, the pheromone concentration matrix of the ant colony optimization algorithm in the flat ground navigation strategy sub-model is initialized in a biased manner, so that the grid nodes located in the direction indicated by the first path geometry information obtain the updated initial pheromone concentration.

11. The mobile robot path optimization system supporting track mode switching as described in claim 10, characterized in that, The construction process of the switching prediction sub-model also includes: When the predicted next target scenario is an orbital scenario, the path optimization feedback data output by the current flat terrain navigation strategy sub-model is received; the path optimization feedback data includes at least the key node sequence or curvature distribution information that characterizes the geometric smoothness of the path, generated during the iterative solution of the flat terrain scenario. Based on the path optimization feedback data, adjust the improved A in the orbital navigation strategy sub-model. The gravitational potential adjustment coefficient corresponding to the potential field heuristic term on which the improved algorithm heuristic function depends; When the mobile robot actually runs to the coordinates of the transition point, the central scheduler switches the currently active navigation strategy sub-model to the target navigation strategy sub-model that has completed data and parameter preloading according to the navigation strategy identifier. The target navigation strategy sub-model uses the pre-transformed initial potential field dataset and pre-loaded optimization parameters to generate continuous control commands for the new scenario.

12. A method for optimizing the path of a mobile robot supporting track mode switching, implemented based on the mobile robot path optimization system supporting track mode switching as described in any one of claims 1-11, characterized in that, include: The operating scenario and real-time load status of the mobile robot are comprehensively encoded to generate navigation task seeds; Based on the navigation task seed, the corresponding navigation strategy model is selected and combined with the preset spatiotemporal reservation table to generate control instructions for the executable trajectory, and an instruction readiness indication is output to the central scheduler. The navigation strategy model includes a track navigation strategy sub-model, a flat ground navigation strategy sub-model, and a switching prediction sub-model. The spatiotemporal reservation table is used to calculate a continuous value reservation weight associated with the probability of multi-vehicle conflict and the remaining safe passage time of the path node for each path node based on the three-dimensional coordinates or two-dimensional grid coordinates of each path node in the executable trajectory and its corresponding time window, so as to optimize the selected track navigation strategy sub-model or flat ground navigation strategy sub-model for multi-vehicle collaborative conflict avoidance and resource scheduling. Upon receiving the instruction ready indication, the control instruction for the executable trajectory is sent to the corresponding track or flat ground navigation control loop.

Images