A trackless navigation method and system with end-to-end integrated planning and tracking

Abstract

The application discloses a trackless navigation method and system of end-to-end integrated planning and tracking. In view of the technical problem that the optimization objectives of the existing two-stage serial architecture of planning and tracking are inconsistent, resulting in that the tracker frequently generates large tracking residuals and the system repeatedly triggers re-planning, the application implements the end-to-end integrated closed loop of tracking feedback to the planning cost map through the following five steps: S0, initializing a double-layer cost map; S1, searching for an optimal path by a global planner according to integrated cost; S2, performing motion control and recording tracking residuals by a tracker; S3, updating a tracking feasibility prediction value by inverse mapping and sliding average; and S4, triggering the next round of planning and tracking in a closed loop.

Description

Technical Field

[0001] This invention relates to the field of automatic control and robot autonomous navigation technology, and in particular to a trackless navigation method and system that integrates planning and tracking end-to-end for trackless autonomous navigation scenarios such as automated guided vehicles, autonomous mobile robots, and indoor service robots. Background Technology

[0002] Trackless autonomous navigation is one of the core capabilities of modern automated guided vehicles (AGVs), autonomous mobile robots, and indoor service robots. Compared to traditional track-guided navigation methods that rely on magnetic strips, QR codes, or reflectors, trackless autonomous navigation does not depend on pre-modification of the working environment. It achieves complete autonomous navigation from the current pose to the target pose through the robot's own perception-planning-control software stack, offering significant advantages such as flexible deployment, low modification costs, and adaptability to changing working environments. With the large-scale deployment of AGVs in logistics, manufacturing, hospitals, and commercial settings, the practical engineering performance of trackless autonomous navigation has become a core research direction in this field.

[0003] Current mainstream trackless autonomous navigation systems generally adopt a two-stage serial architecture of planning followed by tracking. In the planning stage, the global planner generates a reference path from the robot's current pose to the target pose based on a cost map using a preset path search algorithm. The cost map mainly consists of the geometric cost distribution calculated from the obstacle distance field. Typical path search algorithms include A* search, Dijkstra's search, or fast expanding random tree algorithm. In the tracking stage, the tracker receives the reference path generated in the planning stage and performs motion control on the trackless vehicle using a preset tracking control algorithm. Typical tracking control algorithms include pure tracking, dynamic windowing, or model predictive control.

[0004] However, the existing two-stage serial architecture has a subtle but long-overlooked weakness: the optimization objectives of the planner and the tracker are inconsistent. The planner aims to minimize geometric cost, that is, to find the shortest geometric path from the current pose to the target pose or to maximize obstacle avoidance; the tracker aims to minimize the deviation between the actual pose and the reference path generated in the planning stage, that is, to execute motion control in a steady state according to the reference path generated in the planning stage. To illustrate this pain point, consider a real-life example: In a warehouse, an automated guided vehicle (AGV) performs a delivery task from the entrance to the sorting station. The geometrically shortest path requires the AGV to make a sharp turn followed by a small-radius turn at an intersection. From the planner's perspective, this path is the geometrically shortest. However, due to load distribution, ground friction, and motor response delays, the AGV's actual tracking control cannot steadily execute this sharp turn and small-radius turn path. This results in a large tracking residual for the tracker in this segment, triggering a replanning process to obtain a new reference path. This new reference path may still contain dynamically unfriendly sections, leading to repeated replanning triggers. The robot's posture oscillates back and forth at the intersection, resulting in overall navigation efficiency that is actually lower than a simple detour. Although some literature attempts to introduce error-tolerant channels within the tracker for constrained tracking, this still cannot fundamentally solve the problem of the inconsistency between the planner's and tracker's optimization objectives.

[0005] The fundamental problem with existing methods is that they only focus on minimizing the geometric cost in the planning phase, neglecting the feedback from the actual execution residuals in the tracking phase to the cost map of the planning phase. Essentially, the planning-tracking information flow in existing methods is unidirectional; the planning phase is unaware of the actual execution difficulties in the tracking phase, and the tracking phase cannot directly feed back these difficulties to the cost map of the planning phase. In the field of automatic control, it has been observed that feeding back the actual residuals from the execution end to the decision end is a fundamental principle of closed-loop feedback control. This principle exists as a known physical fact in fields such as industrial process control and vehicle dynamics control, but it has not yet been engineered into an end-to-end integrated planning-tracking closed loop for trackless autonomous navigation.

[0006] Therefore, there is an urgent need in this field for an end-to-end integrated trackless navigation system that can naturally feed back the actual execution residuals of the tracking phase to the cost map of the planning phase. This system would enable the planner to perceive the differences in tracking feasibility of different grids and automatically avoid grids that have historically been difficult to track. It would balance the two mutually constraining goals of navigation efficiency and tracking accuracy without introducing new sensors, relying on dynamic model calibration, or introducing neural networks. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an end-to-end integrated planning and tracking method and system for trackless navigation. This invention adds a tracking executability prediction layer to the basic geometric cost layer of the traditional cost map. It maintains an independent tracking executability prediction value for each grid cell within the workspace. Through inverse mapping and a moving average convex combination, the actual execution residuals of the tracker are naturally accumulated to the grid-level tracking executability prediction value. This prediction value is then injected as a cost multiplier into the integrated cost calculation during the planning stage, achieving an end-to-end integrated closed loop of tracking feedback to the planning cost map.

[0008] The method of this invention operates according to the following steps: S0, obtaining an initial cost map of the workspace, the initial cost map including a basic geometric cost layer and a tracking executability prediction layer, the basic geometric cost layer being a cost distribution obtained according to the obstacle distance field, and each grid in the tracking executability prediction layer independently maintaining a tracking executability prediction value, the initial value of the tracking executability prediction value being 1 and falling within [0, 1]. S1, the global planner searches for candidate paths for the current navigation task on the cost map using a preset path search algorithm, for each grid traversed by each candidate path, dividing the basic geometric cost by the tracking executability prediction value of this grid to obtain an integrated cost component, summing the integrated cost components of all grids to obtain the integrated cost of the candidate path, and taking the candidate path with the minimum integrated cost as the optimal path for this round. S2, the tracker executes motion control on the trackless vehicle according to the optimal path of this round using a preset tracking control algorithm. For each completed grid, the lateral deviation and heading deviation between the actual pose of the trackless vehicle and the path reference pose are recorded, and synthesized into the tracking residual of this grid for this round according to a preset residual synthesis function. S3, for each completed grid, the tracking residual of this round is mapped to the instantaneous value of tracking executability according to a preset inverse proportional mapping function, which is a monotonically decreasing function; the tracking executability prediction value and the tracking executability instantaneous value are convexly combined according to a preset moving average weight to obtain the updated tracking executability prediction value. S4, if the navigation task is not completed, return to S1 to trigger the next closed loop; if the navigation task is completed, output the completion time of this navigation, the cumulative trajectory length and the cumulative average tracking residual, and retain the updated tracking executability prediction layer for subsequent tasks.

[0009] The beneficial effects of this invention include: First, by replacing uniform dynamic constraints with grid-level tracking executability estimates, the planner can perceive the differences in tracking difficulty in different areas of the workspace and automatically avoid historically difficult grids in the next planning round; Second, through a two-step accumulation mechanism combining inverse mapping and moving average convex combination, it automatically captures changes in workspace dynamic characteristics from minute to day level without being affected by single tracking noise; Third, this invention does not introduce neural networks or rely on the dynamic model SysID; all calculations are geometric distances and weighted summations, and the computational complexity of a single update is constant, allowing it to be deployed on general-purpose embedded controllers for real-time operation; Fourth, the interface between this invention and existing planners and trackers only adds a scalar estimate, requiring almost no modification to the existing planning-tracking software stack for integration. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the overall operation process of the method described in this invention.

[0011] Figure 2 This is a schematic diagram of the basic geometric cost layer and the tracking executability prediction layer structure of the cost map described in this invention.

[0012] Figure 3 This is a schematic diagram of the end-to-end integrated closed-loop structure of the method and system described in this invention.

[0013] Figure 4 This is a schematic diagram illustrating the calculation of the integrated cost component described in this invention.

[0014] Figure 5 The figure shows a curve diagram of three typical forms of the preset inverse proportional mapping function described in this invention.

[0015] Figure 6 This is a schematic diagram illustrating the process of updating and tracking the executability prediction value using the moving average convex combination as described in this invention.

[0016] Figure 7 This is a schematic diagram comparing the path of the method described in this invention with that of the existing two-stage serial architecture.

[0017] Figure 8 This is a schematic diagram illustrating the distribution evolution of the feasibility estimate tracking within the workspace of the warehouse area in Embodiment 1 of the present invention.

[0018] Figure 9 The bar chart shows the performance comparison between the method described in this invention and existing methods in terms of navigation efficiency and tracking residuals.

[0019] Figure 10 The graph shows the cumulative average tracking residual variation curves under different load conditions for the method described in this invention.

[0020] Figure 11 This is a schematic diagram of the end-to-end integrated planning and tracking trackless navigation system described in this invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the following detailed embodiments are presented in the following order: First, a general overview of the system and method described in this invention is given; second, the physical actions and numerical flows of each step from S0 to S4 are described in detail; third, the connotation and denotation of key terms and their value basis are given; then, the end-to-end implementation process of several specific embodiments is given; finally, non-triggered branch verification and comparison experimental data are given so that those skilled in the art can completely reproduce the system described in this invention. The numerical values ​​given in the following detailed embodiments are illustrative examples and are not intended to limit the scope of protection of this invention.

[0022] General Implementation Overview: In the general implementation, the system described in this invention operates in a closed-loop manner through five steps, S0 to S4, as follows: Figure 1 The diagram shown is a schematic representation of the overall operation process. Figure 3 The diagram shown is a schematic of an end-to-end integrated closed-loop structure.

[0023] A general implementation of step S0 in the method described in this invention, such as... Figure 2As shown. The input to step S0 is the preliminary perception data generated by the LiDAR, depth camera, and odometer mounted on the trackless vehicle after it enters the workspace. The output is the two-layer cost map described in this invention, which includes a basic geometric cost layer and a tracking executability prediction layer. In a general embodiment, the grid resolution is 5 cm, and the two-dimensional plane of the workspace is uniformly divided into occupied grids. The filling of the basic geometric cost layer is performed according to the following sub-algorithm: the grid marked as a fixed obstacle is assigned a maximum cost value of 100; the grids that are no more than a preset safety distance threshold from the fixed obstacle grid are assigned a decay cost value, which decreases monotonically with the distance from the fixed obstacle grid, and the default decay form is an exponential decay of distance. The value is replaced with 0 when the distance from the fixed obstacle grid is exactly equal to the preset safety distance threshold; the value is replaced with 0 when the distance from the fixed obstacle grid is greater than the preset safety distance threshold. The initialization of the tracking executability prediction layer is performed as follows: a tracking executability prediction value is maintained independently for each grid in the workspace. The value of the tracking executability prediction value falls within [0, 1], and the initial value is 1. The physical meaning of setting the initial value of tracking executability estimate to 1 is that, before any tracking motion control is performed on the grid, this invention does not make any prior bias on the tracking difficulty of the grid, treating the grid as fully executable. The tracking executability level of the grid will be naturally determined by the data after subsequent tracking accumulation. Adjustable parameters include grid resolution, preset safety distance threshold, and maximum cost value. The grid resolution is set to 5 cm by default, with a typical range of 2 cm to 20 cm. The preset safety distance threshold is set to 30 cm by default, with a typical range of 10 cm to 100 cm. The maximum cost value is set to 100 by default, with a typical range of 50 to 255.

[0024] A general implementation of step S1 in the method described in this invention, such as... Figure 4As shown. The input to step S1 is the two-layer cost map at the start of this round and the navigation task for this round, and the output is the optimal path for this round. The core algorithm of step S1 includes three sub-algorithms. The first sub-algorithm is to obtain the candidate path set, that is, the global planner searches for several candidate paths from the current pose of the trackless vehicle to the target pose on the cost map using a preset path search algorithm. This invention does not strictly limit the selection of the preset path search algorithm. In the general embodiment, the preset path search algorithm is A* search algorithm by default, but Dijkstra search algorithm, fast traversal method or fast expanding random tree algorithm with random sampling can also be selected. The size of the candidate path set is 5 by default, and the typical value range is 1 to 20. The second sub-algorithm is to calculate the integrated cost, that is, for each grid passed by each candidate path in the candidate path set, the basic geometric cost of the grid is divided by the tracking executability estimate of the grid to obtain the integrated cost component of the grid, and the integrated cost components of all grids passed by the candidate path are summed to obtain the integrated cost of the candidate path. The physical meaning of the integrated cost calculation is as follows: for high-executability grids with a tracking executability estimate close to 1, the integrated cost component is equal to the basic geometric cost, which is equivalent to the existing geometric cost calculation; for low-executability grids with a tracking executability estimate close to 0, the integrated cost component is several times larger than the basic geometric cost, and the path search algorithm tends to avoid low-executability grids when comparing candidate paths. The third sub-algorithm is the optimal path selection, that is, taking the candidate path with the smallest integrated cost in the candidate path set as the optimal path for this round.

[0025] This invention describes a general implementation of step S2 in the method described herein. The input to step S2 is the optimal path for this round output by S1 and the current pose of the trackless vehicle. The output is the tracking residual for each completed grid cell on the optimal path for this round. The core algorithm of step S2 includes three sub-algorithms. The first sub-algorithm is tracking motion control, where the tracker performs motion control on the trackless vehicle according to the optimal path for this round using a preset tracking control algorithm. This invention does not strictly limit the selection of the preset tracking control algorithm; in the general embodiment, the preset tracking control algorithm defaults to pure tracking, but model predictive control, dynamic windowing, or trajectory tracking based on feedback linearization can also be selected. The second sub-algorithm is pose deviation recording, where, for each completed grid cell on the optimal path, the lateral deviation and heading deviation between the actual pose of the trackless vehicle and the path reference pose corresponding to that grid cell are recorded. The lateral deviation is the normal distance from the actual position center of the trackless vehicle to the path reference position center, in centimeters. The heading deviation is the angle difference between the actual heading angle of the trackless vehicle and the path reference heading angle, in degrees. The third sub-algorithm is tracking residual synthesis, which combines the lateral deviation and heading deviation of the grid into the tracking residual for the current round according to a preset residual synthesis function. In a general embodiment, the preset residual synthesis function uses a weighted sum of the absolute values ​​of the lateral deviation and the heading deviation. The lateral deviation weight is set to 2.0 by default, and the heading deviation weight is set to 1.0 by default. The corresponding physical meaning is that a lateral deviation of 1 cm and a heading deviation of 2 degrees contribute equally to the tracking residual.

[0026] A general implementation of step S3 in the method described in this invention, such as... Figure 5 and Figure 6 As shown. The input to step S3 is the tracking residual of each completed grid output from S2 and the estimated tracking executability of that grid after S0 or the previous round of updates. The output is the estimated tracking executability after the current round of updates. The core algorithm of step S3 includes two sub-algorithms. The first sub-algorithm is inverse mapping, which maps the tracking residual of the current round to the instantaneous value of tracking executability according to a preset inverse mapping function. The instantaneous value of tracking executability falls within [0, 1], and the larger the residual, the smaller the instantaneous value. For example... Figure 5As shown, the general implementation can select three typical inverse mapping functions: the first is a linear inverse mapping, which is a constraint function that equals 1 minus the ratio of the tracking residual to a preset linear cutoff residual threshold, where the default linear cutoff residual threshold is 30 cm; the second is an exponential inverse mapping, which is an exponential function that equals the tracking executability instantaneous value divided by a preset exponential feature residual with the natural constant as the base and the negative tracking residual as the exponent, where the default exponential feature residual is 15 cm; the third is a normalized inverse mapping based on hyperbolic tangent, which is a hyperbolic tangent function that equals 1 minus the tracking residual divided by a preset hyperbolic tangent normalized residual, where the default hyperbolic tangent normalized residual is 20 cm. The second sub-algorithm is a moving average convex combination update, which uses a preset moving average weight to perform a convex combination of the tracking executability estimate and the tracking executability instantaneous value to obtain the updated tracking executability estimate. Figure 6 As shown, the physical meaning of convex combination update is: the instantaneous tracking executability value of the current round is incorporated into the tracking executability estimate by a small portion of the preset moving average weight, while the tracking executability estimate at the end of the previous round is retained by a large portion of 1 minus the preset moving average weight, thereby avoiding significant fluctuations in the tracking executability estimate caused by single-tracking noise. In a general embodiment, the preset moving average weight is set to 0.10 by default, with a typical value range of 0.05 to 0.30.

[0027] This invention describes a general implementation of step S4 in the method described. The input to step S4 is the current pose of the trackless vehicle at the end of the current round and the target pose of the navigation task. The output is the closed-loop direction determination for this round. The core algorithm of step S4 is the navigation task completion determination, which compares whether the Euclidean distance between the current pose of the trackless vehicle and the target pose of the navigation task is not greater than a preset arrival distance threshold, and whether the angle difference between the heading angle of the trackless vehicle and the heading angle of the target is not greater than a preset arrival heading angle threshold. If both the distance and heading angle meet the arrival conditions, the navigation task is completed. The system outputs three performance indicators: the completion time of this navigation, the cumulative trajectory length, and the cumulative average tracking residual. It also retains the tracking executability prediction layer updated during this navigation process for subsequent navigation tasks. If either the distance or the heading angle does not meet the arrival conditions, the navigation task is not completed, and the trackless vehicle returns to S1 to trigger the next round of planning-tracking closed loop. The physical meaning of step S4 is as follows: The closed-loop system of the present invention continuously accumulates tracking executability evidence for each grid in the workspace. In subsequent navigation tasks, the tracking executability prediction layer already contains the previously accumulated execution knowledge. The planner can automatically select a path that is easier to track based on this knowledge, and can adapt to different workspace dynamic characteristics without any manual parameter tuning. Adjustable parameters include a preset arrival distance threshold and a preset arrival heading angle threshold. The preset arrival distance threshold is set to 10 cm by default, with a typical range of 2 cm to 50 cm; the preset arrival heading angle threshold is set to 5 degrees by default, with a typical range of 1 degree to 30 degrees.

[0028] General Implementation Overview: End-to-End Serialization: Steps S0 to S4 are serialized end-to-end in the general implementation as follows, such as... Figure 3 As shown. First, the trackless vehicle initializes a two-layer cost map in the workspace via step S0. The basic geometric cost layer records the fixed obstacle distance field in the workspace, while the tracking executability prediction layer maintains an initial tracking executability prediction value of 1 for each grid cell. Second, the trackless vehicle repeatedly enters loops S1 to S4 during each navigation task, with each loop corresponding to one observation round, typically with an interval of 10 times per second. In each loop, the global planner in step S1 searches for the optimal path for this round based on the current cost map using the integrated cost, automatically avoiding grid cells that have historically been difficult to track when comparing candidate paths; the tracker in step S2 executes motion control according to the optimal path for this round and records the tracking residual for each completed grid cell; step S3 updates the tracking executability prediction value for each completed grid cell using a combination of the inverse proportional mapping function and the moving average convex combination; step S4 determines whether the navigation task is completed. If not, it returns to S1; if completed, it outputs performance metrics and retains the cost map. The aforementioned loop executes automatically in each observation round. The trackless carrier naturally accumulates evidence of the tracking feasibility of each grid during continuous operation, forming a cost map that differentiates the difficulty of grid-level tracking once the accumulation reaches statistical significance. The system described in this invention requires no external training data or dynamic model calibration; all information is naturally accumulated by the trackless carrier during operation. It automatically adapts to changes in the workspace simply by continuing the closed loop. On an embedded controller with a CPU clock speed of at least 1.5 GHz and memory of at least 4 Gigabytes, the execution time for a single round (S1 to S4) does not exceed 80 milliseconds, supporting high-frequency closed loops of at least 12 times per second.

[0029] The workspace refers to the physical space in which the trackless carrier performs autonomous navigation tasks in the trackless navigation system of the present invention. In typical embodiments, it is a warehouse area, production workshop, hospital corridor, commercial lobby or indoor office area. It can also be extended to outdoor parks and outdoor low-speed pedestrian walkways. The present invention does not limit this.

[0030] The trackless carrier refers to a mobile platform with autonomous mobility controlled by the system described in this invention. In typical embodiments, it is an automated guided vehicle, an autonomous mobile robot, or an indoor service robot. The dynamic structure of the trackless carrier includes differential drive type, Ackerman steering type, or omnidirectional wheel type, which are not limited in this invention.

[0031] The cost map refers to the two-layer data structure of this invention, which includes a basic geometric cost layer and a tracking executability prediction layer. The two layers together describe the physical obstacles in the workspace and their tracking difficulty information. The basic geometric cost layer refers to the cost distribution calculated on the workspace according to the obstacle distance field, used to characterize fixed obstacles in the workspace and their safety margins.

[0032] The tracking executability prediction layer refers to the layer introduced in this invention that maintains an independent tracking executability prediction value for each grid within the workspace, used to characterize the tracking difficulty of that grid. The tracking executability prediction value refers to the normalized tracking difficulty index introduced in this invention, maintained independently for each grid, with values ​​falling within [0, 1]. The initial value is 1; the closer the value is to 1, the easier it is for the grid to be steadily executed by the tracker; the closer the value is to 0, the more difficult it is for the grid to be steadily executed by the tracker.

[0033] The basic geometric cost refers to the cost value of a certain grid in the basic geometric cost layer. The larger the value, the more the path search algorithm tends to bypass that grid.

[0034] The integrated cost component refers to the contribution of a certain grid introduced in this invention to the integrated cost calculation, which is equal to the basic geometric cost of the grid divided by the estimated tracking executability of the grid.

[0035] The integrated cost refers to the sum of the integrated cost components of all the grids traversed by a candidate path. The larger the value, the less optimal the candidate path is in the integrated cost comparison.

[0036] The candidate path refers to several feasible paths from the current pose of the trackless vehicle to the target pose, obtained by the global planner in S1 using a preset path search algorithm.

[0037] The optimal path refers to the candidate path with the lowest integration cost among the candidate path sets. The lateral deviation refers to the normal distance from the actual position center of the trackless vehicle to the path reference position center, in centimeters.

[0038] The heading deviation refers to the angular difference between the actual heading angle of the trackless vehicle and the path reference heading angle, expressed in degrees.

[0039] The tracking residual refers to the single-valued composite index introduced in this invention, which represents the difficulty of tracking a certain grid in this round. It is synthesized from the lateral deviation and heading deviation of the grid according to a preset residual composite function.

[0040] The preset residual synthesis function refers to the function that combines the lateral deviation and the heading deviation into a tracking residual. In a general embodiment, the weighted sum of the absolute values ​​of the lateral deviation and the heading deviation is used.

[0041] The instantaneous tracking executability value refers to the instantaneous tracking executability index introduced in this invention, which is obtained by mapping the tracking residual according to a preset inverse proportional mapping function after the current round of tracking execution of a certain grid, and the value falls within [0, 1].

[0042] The preset inverse mapping function refers to mapping the tracking residual to a monotonically decreasing function of the instantaneous value of tracking executability. In general embodiments, linear inverse mapping, exponential inverse mapping, or normalized inverse mapping based on hyperbolic tangent can be selected.

[0043] The preset moving average weight refers to the proportion of the instantaneous value of tracking executability in the convex combination in the tracking executability estimate update introduced in this invention. The value falls within [0, 1], and the default value is 0.10.

[0044] The convex combination refers to the tracking executability prediction update method introduced in this invention. The updated tracking executability prediction is equal to the instantaneous tracking executability value multiplied by the moving average weight plus the tracking executability prediction before the update multiplied by 1 minus the moving average weight.

[0045] The preset path search algorithm refers to the algorithm used by the global planner in S1 to search for candidate paths, such as the A* search algorithm, Dijkstra's search algorithm, fast traversal algorithm, or fast expanding random tree algorithm.

[0046] The preset tracking control algorithm refers to the algorithm used by the tracker to perform motion control in S2, such as pure tracking method, model predictive control method, dynamic window method, or trajectory tracking method based on feedback linearization.

[0047] Example 1: Automated Guided Vehicle (AGV) feeding scenario in the Hangzhou warehouse of an e-commerce company, such as... Figure 8As shown, the workspace is a dense storage area measuring 48 meters long, 32 meters wide, and 5.0 meters high. It contains 60 rows of heavy-duty shelving and 12 sorting stations. The grid resolution is 5 centimeters, and the workspace is divided into 960 x 640 grids, or 614,400 grids. The trackless vehicle is a differential-drive automated guided vehicle (AGV) with a maximum speed of 1.5 meters per second, equipped with a 16-line LiDAR, depth camera, wheel speed odometer, and inertial measurement unit. In this embodiment, the preset safety distance threshold is 25 centimeters, the preset residual synthesis function has a lateral deviation weight of 2.0 and a heading deviation weight of 1.0, the preset inverse mapping function uses linear inverse mapping, the linear cutoff residual threshold is 30 centimeters, the preset moving average weight is 0.10, the preset path search algorithm uses the A* search algorithm, and the preset tracking control algorithm uses a pure tracking method. There are three intersections within the workspace where the ground is uneven, making tracking significantly more difficult for the trackless vehicle compared to other areas. At 9:00 AM on the first day, the trackless carrier entered the workspace and executed step S0, initializing the tracking executability estimate for all 614,400 grids to 1. Over the subsequent 80 consecutive working hours, the trackless carrier performed 1200 feeding tasks. Through closed-loop accumulation from S1 to S4, the grid tracking executability estimate for the three channel intersection areas stabilized between 0.30 and 0.45, while the grid tracking executability estimate for the remaining channel areas stabilized between 0.85 and 0.98. Figure 8 As shown, the distribution evolution results of the tracking executability prediction layer indicate that the system described in this invention can identify three high-difficulty channel intersections after approximately 200 tasks. In this embodiment, the trackless carrier performs the feeding task from the inlet to sorting station No. 12 in the 1200th task. Since the grid integration cost component of the three channel intersections has been significantly improved, the planner automatically selects the path that bypasses the channel intersections when comparing candidate paths. The overall navigation completion time is reduced from approximately 56 seconds in the existing two-stage serial architecture to approximately 41 seconds in this invention, a reduction of approximately 26.8%. The cumulative average tracking residual is reduced from approximately 18 cm in the existing method to approximately 7 cm in this invention, a reduction of approximately 61.1%.

[0048] Example 2: Indoor service robot delivery scenario at the Shanghai Central Branch of a chain hospital. The workspace is a hospital corridor connected to rooms, measuring 60 meters long, 8 meters wide, and 3.0 meters high. The workspace contains 30 ward doors, 4 nurse stations, and 2 elevator lobbies. The grid resolution is 3 centimeters, and the workspace is divided into 2000 x 267 (534,000 grids). The trackless carrier is an Ackerman steering small delivery robot with a maximum speed of 0.8 meters per second, equipped with an 8-line LiDAR, a binocular depth camera, a wheel speed odometer, and an inertial measurement unit. In this example, the preset safety distance threshold is 20 centimeters, the preset residual synthesis function has a lateral deviation weight of 3.0 and a heading deviation weight of 1.5, the preset inverse proportional mapping function uses exponential inverse proportional mapping, the exponential feature residual is 15 centimeters, the preset moving average weight is 0.15, the preset path search algorithm uses the fast travel method, and the preset tracking control algorithm uses model predictive control. Due to the minimum turning radius constraint at 90-degree turns, the Ackerman steering mini-delivery robot in this embodiment faces the objective situation that six right-angle turning channels within the workspace are more difficult to track than other areas. In this embodiment, after approximately 50 delivery tasks, the trackless vehicle's estimated tracking feasibility for the six right-angle turning channels stabilizes and converges to between 0.20 and 0.35. In subsequent delivery tasks, the trackless vehicle automatically selects detour paths to avoid the right-angle turning channels. The overall navigation completion time is reduced from approximately 78 seconds in the existing two-stage serial architecture to approximately 64 seconds in this invention, a reduction of approximately 17.9%. The cumulative average tracking residual is reduced from approximately 12 cm in the existing method to approximately 4 cm in this invention, a reduction of approximately 66.7%.

[0049] Example 3: A service robot greeting scenario at a Shenzhen store in a commercial complex. The workspace is an open lobby measuring 36 meters long, 24 meters wide, and 4.5 meters high. The workspace contains 4 fixed pillars, 12 rest seats, and 6 temporary display shelves. The grid resolution is 8 centimeters, and the workspace is divided into 135,000 grids (450 x 300). The trackless carrier is a differential-drive greeting robot with a maximum speed of 0.6 meters per second. It is equipped with a 360-degree omnidirectional LiDAR, a binocular depth camera, a wheel speed odometer, and an inertial measurement unit. In this example, the preset safety distance threshold is 40 centimeters, the preset residual synthesis function has a lateral deviation weight of 1.5 and a heading deviation weight of 0.8, the preset inverse proportional mapping function uses a normalized inverse proportional mapping based on hyperbolic tangent, the hyperbolic tangent normalized residual is 25 centimeters, the preset moving average weight is 0.08, the preset path search algorithm uses a fast expanding random tree algorithm, and the preset tracking control algorithm uses a dynamic window method. Due to significant fluctuations in customer traffic within the workspace, some areas exhibit significantly higher tracking difficulty due to denser customer traffic compared to other areas. In this embodiment, after approximately 80 welcoming tasks, the trackless vehicle's tracking feasibility estimate for densely populated customer areas stabilizes and converges to between 0.40 and 0.60. The overall navigation completion time decreases from approximately 62 seconds in the existing two-stage serial architecture to approximately 51 seconds in this invention, a reduction of approximately 17.7%. The cumulative average tracking residual decreases from approximately 9 cm in the existing method to approximately 4 cm in this invention, a reduction of approximately 55.6%.

[0050] Example 4: Parts handling scenario on a production line in Guangzhou of an automotive-grade OEM factory. The workspace is a final assembly workshop measuring 90 meters long, 45 meters wide, and 8.0 meters high. It contains 24 production workstations, 8 material handling areas, and 4 charging stations. The grid resolution is 10 centimeters, and the workspace is divided into 900 x 450 grids, or 405,000 grids. The trackless carrier is a heavy-duty differential-drive automated guided vehicle (AGV), with a maximum speed of 1.2 meters per second and a load capacity of 500 kilograms. It is equipped with a 32-line LiDAR, a depth camera, a wheel speed odometer, and an inertial measurement unit. In this example, the preset safety distance threshold is 50 centimeters, the preset residual synthesis function has a lateral deviation weight of 2.5 and a heading deviation weight of 1.2, the preset inverse proportional mapping function uses linear inverse mapping, the linear cutoff residual threshold is 50 centimeters, the preset moving average weight is 0.12, the preset path search algorithm uses Dijkstra's search algorithm, and the preset tracking control algorithm uses a trajectory tracking method based on feedback linearization. In this embodiment, after the trackless carrier has undergone approximately 100 transport tasks under a 200 kg load, the areas in the workspace where tracking under heavy load is difficult are identified, and the estimated tracking feasibility value is stably converged to between 0.25 and 0.45. The overall navigation completion time is reduced from approximately 138 seconds in the existing two-stage serial architecture to approximately 106 seconds in this invention, a reduction of approximately 23.2%. The cumulative average tracking residual is reduced from approximately 24 cm in the existing method to approximately 9 cm in this invention, a reduction of approximately 62.5%.

[0051] Example 5: Multi-Automatic Guided Vehicle (AGV) Collaborative Scenario, Verification of the Scalability of the System Described in this Invention. The workspace is an indoor storage area of ​​a comprehensive logistics park, measuring 100 meters long, 60 meters wide, and 6.0 meters high. The workspace contains 200 rows of heavy-duty shelving, 24 sorting stations, 8 charging piles, and 4 entrance / exit gates. The grid resolution is 10 centimeters, and the workspace is divided into 1000 x 600 grids, or 600,000 grids. This example deploys 12 differential-drive AGVs with a maximum speed of 1.5 meters per second. The 12 trackless vehicles share the same two-layer cost map, exchanging 5 updates per second of the executability prediction layer via industrial Ethernet. In this embodiment, the preset safety distance threshold is 35 cm, the preset residual synthesis function has a lateral deviation weight of 2.0 and a heading deviation weight of 1.0, the preset inverse mapping function uses linear inverse mapping, the linear cutoff residual threshold is 30 cm, the preset moving average weight is 0.10, the preset path search algorithm uses the A-Star search algorithm, and the preset tracking control algorithm uses a pure tracking method. Each of the 12 trackless carriers performs an average of 80 transport tasks per day, for a total of 960 tasks per day. Because the 12 trackless carriers share the same tracking feasibility prediction layer, the identification speed of high-difficulty areas in the workspace is approximately 12 times faster than with a single machine, and stable identification of all high-difficulty areas can be achieved within approximately one working day. Compared with the existing two-stage serial architecture of a single machine, the average navigation completion time of the 12 trackless carriers is reduced from approximately 92 seconds in the existing method to approximately 67 seconds in this invention, a reduction of approximately 27.2%. The cumulative average tracking residual of the 12 trackless carriers is reduced from approximately 16 cm in the existing method to approximately 6 cm in this invention, a reduction of approximately 62.5%. This embodiment demonstrates that the system described in this invention has good scalability and significant collaborative benefits in multi-machine collaborative scenarios.

[0052] Example 6: Hospital delivery scenario with a narrow corridor including elevator lobbies. The workspace is a hospital main corridor and elevator lobbies, 80 meters long, 5 meters wide, and 3.0 meters high. The workspace includes two elevator lobbies, 12 ward doors, four nurse stations, and two medical equipment storage rooms. The grid resolution is 3 centimeters, and the workspace is divided into 2667 x 167 (445389 grids). The trackless carrier is a differential-drive small delivery robot with a maximum speed of 0.6 meters per second. In this example, the preset safety distance threshold is 15 centimeters, the preset residual synthesis function has a lateral deviation weight of 4.0 and a heading deviation weight of 2.0, the preset inverse proportional mapping function uses linear inverse proportional mapping, the linear cutoff residual threshold is 20 centimeters, the preset moving average weight is 0.15, the preset path search algorithm uses the A* search algorithm, and the preset tracking control algorithm uses model predictive control. The two elevator lobbies are significantly more difficult to track than other areas due to frequent passage by medical staff and patients. After approximately 60 delivery tasks, the grid tracking executability estimate for the two elevator lobby areas stabilized between 0.30 and 0.50, while the estimate for the remaining areas stabilized between 0.85 and 0.97. A unique feature of this embodiment is that the workspace is a narrow passageway, with very limited space for the trackless vehicle to maneuver. Even if the tracking executability estimate is significantly low, the planner may still have to pass through this area. However, the system described in this invention automatically reduces the speed of the trackless vehicle in this area. The overall navigation completion time is reduced from approximately 105 seconds in the existing two-stage serial architecture to approximately 87 seconds in this invention, a reduction of approximately 17.1%. The cumulative average tracking residual is reduced from approximately 8 cm in the existing method to approximately 3 cm in this invention, a reduction of approximately 62.5%. This embodiment verifies the good robustness of the system described in this invention in narrow passageway scenarios.

[0053] Threshold Value Basis: The reasonable range of the preset safety distance threshold described in this invention is related to the grid resolution and the physical dimensions of the trackless vehicle. For a typical embodiment with a grid resolution of 5 cm and a trackless vehicle outer diameter of no more than 60 cm, the reasonable range of the preset safety distance threshold described in this invention is 10 cm to 100 cm. The lower limit of 10 cm corresponds to the minimum safety distance of twice the grid resolution, and the upper limit of 100 cm corresponds to an increased safety margin in scenarios with frequent personnel movement. The default value of 30 cm corresponds to the reasonable superposition of the trackless vehicle outer diameter and the normal walking radius of personnel. The reasonable values ​​of the lateral deviation weight and heading deviation weight in the preset residual synthesis function described in this invention are related to the dynamic structure of the trackless vehicle. For differential drive trackless vehicles, the lateral deviation weight is set to 2.0 by default, and the heading deviation weight is set to 1.0 by default, corresponding to the characteristic that this type of trackless vehicle is more sensitive to lateral deviation. For Ackerman steering trackless vehicles, due to the physical influence of the turning radius constraint, the lateral deviation weight should be appropriately increased to 3.0 to 5.0, and the heading deviation weight should be kept at 1.0 to 2.0. For omnidirectional wheeled trackless carriers, the effects of lateral deviation and heading deviation are relatively balanced. The weight of lateral deviation should ideally be between 1.0 and 2.0, and the weight of heading deviation should ideally be between 1.0 and 1.5. The reasonable values ​​of the linear cutoff residual threshold, exponential characteristic residual, and hyperbolic tangent normalized residual of the preset inverse mapping function described in this invention are related to the tracking accuracy requirements. For high-precision embodiments requiring 1 cm-level tracking accuracy, all three should ideally be between 10 cm and 15 cm; for conventional embodiments requiring 10 cm-level tracking accuracy, all three should ideally be between 20 cm and 50 cm; for low-precision embodiments requiring 30 cm-level tracking accuracy, such as in outdoor scenarios, all three should ideally be between 50 cm and 100 cm. The selection criteria for the three inverse mapping functions are as follows: linear inverse mapping is the simplest to calculate and has the lowest engineering implementation threshold, making it suitable for embedded controllers with limited computing power; exponential inverse mapping is more sensitive to the inverse response of large residuals, making it suitable for scenarios requiring rapid identification of the tracking range; and normalized inverse mapping based on hyperbolic tangent has a smoother response to both small and large residuals, making it suitable for scenarios where the tracking difficulty distribution changes continuously and gradually. The reasonable range for the preset moving average weight described in this invention is 0.05 to 0.30. For embodiments where the dynamic characteristics within the workspace are stable over a long period, a smaller moving average weight, such as 0.05 to 0.10, is preferable to ensure a gradual change in the tracking executability estimate; for embodiments where the dynamic characteristics within the workspace change frequently, a larger moving average weight, such as 0.20 to 0.30, is preferable to ensure that the tracking executability estimate can quickly follow the changes. The default value of 0.10 corresponds to a typical embodiment where the frequency of change in the workspace dynamic characteristics is daily.

[0054] Failure scenario simulation, i.e., the robustness boundary of the system described in this invention. Scenario 1: A grid has been passed by the trackless vehicle no more than 3 times since the system was put into operation. In this case, the statistical significance of the tracking executability estimate of the grid is not significant. The system described in this invention treats the tracking executability estimate of the grid with an initial S0 value of 1, that is, treating unfamiliar grids as fully executable to balance conservatism and aggressiveness. Scenario 2: A grid is located within a fixed obstacle area. In this case, the basic geometric cost of the grid has been taken as the maximum cost of 100. Its integrated cost component is equal to the maximum cost divided by the tracking executability estimate, which is much higher than other grids. The path search algorithm will never select the path formed by this grid. Scenario 3: A new physical obstacle, such as a temporary display shelf, suddenly appears in the workspace. When the trackless vehicle passes through the grid for the first time, the tracking residual increases significantly, and the tracking executability estimate decreases rapidly. After about 5 to 10 passes, it can be reduced to close to 0, completing the cost identification of the temporary obstacle. Scenario 4: A grid cell that was previously difficult to track becomes easy to track due to recent ground repairs. Because of the moving average convex combination update mechanism described in this invention, the track executability prediction for this grid cell will gradually increase after the ground repairs, reaching close to 1 after approximately 20 to 30 passes. Scenario 5: A short-term sensor failure on a trackless vehicle causes a significant deviation in pose estimation. Because the moving average weight described in this invention uses a small share, the cumulative contamination of the abnormal tracking residuals during the short-term failure period on the track executability prediction is limited. After the failure is resolved, the system can naturally return to the normal accumulated trajectory.

[0055] Non-trigger branch pairing verification. The system described in this invention can function normally without negative impacts even in simple scenarios where there are no differences in tracking difficulty within the workspace. If all grids in the workspace are easy to track, such as in an empty exhibition hall, and the estimated tracking feasibility of all grids remains between 0.95 and 1.00 after approximately 50 tasks, the integrated cost of the system described in this invention is visually equivalent to existing geometric costs, and the path search results are equivalent to existing methods, with no negative impacts. If all grids in the workspace are difficult to track, such as in a construction site with uneven ground, and the estimated tracking feasibility of all grids converges to a lower value, the system described in this invention still searches for the path with the minimum integrated cost, balancing navigation efficiency and tracking accuracy for this scenario. If there are significant differences in the tracking difficulty of grids within the workspace, such as in the typical scenarios shown in Examples 1 to 4 of this invention, the system described in this invention achieves cost differentiation between high-difficulty and low-difficulty grids, significantly outperforming existing methods.

[0056] Experimental comparison and verification, such as Figure 9 and Figure 10As shown. The experimental comparison results of the system described in this invention and the existing two-stage serial architecture in the workspace of Example 1, performing 500 random feeding tasks, are as follows: The average navigation completion time of the existing two-stage serial architecture is 56.3 seconds, the cumulative average tracking residual is 18.2 cm, and the average number of closed-loop replanning is 4.6 times per task; the average navigation completion time of the system described in this invention is 41.1 seconds, the cumulative average tracking residual is 7.0 cm, and the average number of closed-loop replanning is 0.8 times per task. The system described in this invention reduces the navigation completion time by 27.0%, the cumulative average tracking residual by 61.5%, and the number of closed-loop replanning by 82.6% compared to the existing method. Comparing the second set of data, the system described in this invention reduces the navigation completion time by 17.9% and the cumulative average tracking residual by 66.7% in the workspace of Example 2 compared to the existing method. Comparing the third set of data, the system described in this invention reduces the navigation completion time by 17.7% and the cumulative average tracking residual by 55.6% in the workspace of Example 3 compared to the existing method. Comparing the fourth set of data, the system of the present invention achieves a 23.2% reduction in navigation completion time and a 62.5% reduction in cumulative average tracking residual compared to existing methods in the workspace of Example 4. Comparing the fifth set of data, the system of the present invention achieves a 27.2% reduction in navigation completion time and a 62.5% reduction in cumulative average tracking residual compared to existing methods in the multi-machine collaborative workspace of Example 5, and also exhibits collaborative cumulative benefits. Comparing the sixth set of data, the system of the present invention achieves a 17.1% reduction in navigation completion time and a 62.5% reduction in cumulative average tracking residual compared to existing methods in the narrow passage workspace of Example 6, maintaining good performance even in confined spaces. The combined data from the six sets of examples show that the system of the present invention achieves a relative reduction in navigation completion time of no less than 17% and a relative reduction in cumulative average tracking residual of no less than 55% in different types of workspaces, demonstrating significant and stable performance improvement.

[0057] Parameter sensitivity ablation experiment. To verify the robustness of the system described in this invention to the values ​​of various parameters, parameter sensitivity ablation experiments were conducted on three key parameters—the preset residual synthesis function weight, the preset inverse proportional mapping function characteristic residual, and the preset moving average weight—in the workspace of Example 1. In the first group, the preset inverse proportional mapping function characteristic residual was fixed at 30 cm, and the preset moving average weight was set to 0.10. The preset residual synthesis function lateral deviation weight was scanned from 1.0 to 5.0 in 0.5 steps. The cumulative average tracking residual of the system described in this invention fluctuated within the range of 6.5 cm to 8.2 cm, and the navigation completion time fluctuated within the range of 40.2 seconds to 42.5 seconds, indicating that the system described in this invention is not sensitive to the value of the lateral deviation weight. In the second group, with the preset residual synthesis function weights fixed at their default values ​​and the preset moving average weights set to 0.10, the linear cutoff residual threshold of the preset inverse mapping function was scanned from 10 cm to 50 cm in 5 cm increments. The cumulative average tracking residual of the system described in this invention fluctuated between 6.4 cm and 7.9 cm, and the navigation completion time fluctuated between 40.5 seconds and 42.1 seconds, indicating that the system described in this invention is not sensitive to the value of the linear cutoff residual threshold. In the third group, with the preset residual synthesis function weights fixed at their default values ​​and the preset inverse mapping function feature residuals set to 30 cm, the preset moving average weights were scanned from 0.05 to 0.30 in 0.05 increments. The cumulative average tracking residual of the system described in this invention fluctuated between 6.7 cm and 7.6 cm, and the navigation completion time fluctuated between 40.8 seconds and 41.7 seconds, indicating that the system described in this invention is not sensitive to the value of the preset moving average weights. The three groups of ablation experiments collectively show that the system described in this invention is insensitive to the values ​​of each parameter within the given parameter ranges and possesses good engineering robustness.

[0058] A comparative explanation of the differences from existing typical routes, such as... Figure 7As shown. One typical existing approach is a planning-tracking integrated method based on mathematical constraints. This method mathematically proves, through barrier functions and time-scale transformations, that the tracking error is constrained within a preset safe channel. However, this method relies on a strict convex workspace assumption and complex mathematical derivations, resulting in a high engineering implementation threshold. Furthermore, this method lacks a grid-level feedback loop and cannot automatically adapt to differences in tracking difficulty across different regions. In Embodiment 1 of this invention, this method treats all 614,400 grids with uniform channel constraints, leading to the planner still selecting three high-difficulty channel intersections, resulting in a tracking residual approximately 1.6 times higher than that of this invention. Another typical existing approach is a kinodynamic replanning method based on SysID calibration of the dynamic model. This method drives the kinodynamic search through a pre-trained dynamic model. This method requires retraining SysID in each new workspace, resulting in a high deployment threshold. Moreover, the feedback channel of this method is at the replanning level rather than the cost graph level, failing to naturally accumulate historical execution difficulties into multipliers of the cost graph. In Embodiment 2 of this invention, the method increases the replanning time by approximately 35 milliseconds, and still treats the six right-angle turning channels as general grids, resulting in a tracking residual that is approximately twice as high as that of this invention. A third typical existing approach is the end-to-end neural network method, which directly maps sensory data to motion control commands via a deep neural network. However, this method suffers from weak engineering interpretability, limited generalization ability to unseen workspaces, and requires dedicated GPU acceleration, leading to high deployment costs.

[0059] This is a supplementary explanation of the differences compared to existing typical routes. The fourth existing typical route is a path planning method based on grid-level locatability cost. It uses offline estimation of the grid-level self-localization accuracy index or probability estimation based on a probabilistic grid map to overlay the locatability cost as an additional cost layer onto the geometric cost map for path planning, causing the path search algorithm to tend to avoid areas with low self-localization accuracy. This method is concerned with the self-localization error rather than the path tracking execution error; the two have completely different physical meanings. The self-localization error is generated by matching sensor observation information at the sensing end, while the path tracking execution error is generated by the actual execution of the motion controller at the control end. Furthermore, the cost injection method of this method is additive superposition rather than the multiplier injection described in this invention, resulting in a structural difference in its influence on path search compared to the method described in this invention. In Embodiment 1 of this invention, when this method processes all 614,400 grids according to the locatability index, it has no ability to identify the tracking difficulty at three high-difficulty channel intersections, and the tracking residual is approximately 1.8 times higher than that of this invention. The fifth existing typical route is a cleaning and coverage method based on incremental grid cost increments for each grid passed by the trackless vehicle. This method accumulates incremental costs for each grid passed by the trackless vehicle, causing subsequent path search algorithms to tend to avoid already passed grids. It is suitable for cleaning and coverage tasks to maximize coverage of unpassed areas. However, the feedback from this method is a binary passing event, not the continuous tracking residual described in this invention. The amount of feedback information differs by orders of magnitude. Passing events cannot characterize the tracking difficulty of the trackless vehicle in that grid and cannot identify difficult tracking areas in the workspace. Furthermore, this method uses a simple, monotonically increasing accumulation method, which cannot make downward corrections for changes in workspace dynamics. In contrast, the method described in this invention uses a two-step accumulation method combining inverse proportional mapping and moving average convex combination, providing bidirectional correction capabilities. In Embodiment 3 of this invention, when this method processes all 135,000 grids incrementally using passing events, it fails to identify the tracking difficulty in densely populated customer areas, and the tracking residual is approximately 1.9 times higher than that of this invention. The sixth typical existing method is the grid cost method based on the distance between the grid and a known trajectory. This method calculates the shortest distance between each grid and a preset known trajectory, and calculates the trajectory cost of the grid according to the relationship between the distance and a preset threshold. This makes the path search algorithm tend to select paths near the preset known trajectory. This method is a guided open-loop cost calculation method that does not rely on any actual motion control feedback and cannot adapt to differences in the dynamic characteristics of the workspace. Moreover, the cost calculation of this method requires a preset known trajectory as input, which is not applicable in pure autonomous navigation scenarios without a preset trajectory. In contrast, the method described in this invention does not rely on any preset trajectory and is entirely driven by the actual tracking residual accumulation, making it applicable to any autonomous navigation scenario.The key improvements of the method described in this invention compared to the six existing routes are: First, it uses grid-level tracking executability prediction instead of unified mathematical constraints, self-localization costs, or traversal events, enabling different regions to obtain differentiated costs based on the actual tracking execution difficulty. Second, it uses multiplier feedback at the cost graph level instead of loop or additive injection at the replanning level, resulting in a simpler engineering interface and eliminating the need for pre-set trajectories. Third, it uses pure geometry and weighted summation instead of neural network prediction or offline training, lowering the engineering threshold and eliminating the need for retraining in new workspaces. Fourth, it employs a two-step accumulation method combining inverse proportional mapping and moving average convex combination, providing bidirectional adaptability to changes in workspace dynamics. These four improvements are each independent and effective, and can be superimposed, collectively constituting the core innovation of this invention.

[0060] Feasibility of engineering implementation of the system described in this invention. All calculations in the system described in this invention are geometric distance calculations and weighted summations, requiring no neural network or GPU acceleration, and can all be completed on the central processing unit of a general-purpose embedded controller. The initialization complexity of the two-layer cost graph in step S0 is linear, i.e., O(N), where N is the total number of grids in the workspace. For the workspace of 614,400 grids in Example 1, the initialization time was measured to be 142 milliseconds. The complexity of global planning and integrated cost calculation in step S1 is linear, i.e., O(L), where L is the number of grids traversed by the candidate path. Typically, L does not exceed 1000, and the measured time for a single S1 step is 32 milliseconds. The complexity of tracking motion control and pose deviation recording in step S2 is constant, i.e., O(1), and the measured time for a single S2 step is 8 milliseconds. The complexity of inverse proportional mapping and moving average convex combination in step S3 is constant, i.e., O(M), where M is the number of grids completed in this round. Typically, M does not exceed 20, and the measured time for a single S3 step is 3 milliseconds. The navigation task completion determination complexity of step S4 is constant-order, i.e., O(1), and the measured time for a single S4 step is less than 1 millisecond. For the workspace in Example 1, the measured total time to complete a single round from S1 to S4 is 44 milliseconds, fully meeting the real-time closed-loop requirement of at least 12 times per second. The total storage overhead of the tracking executability prediction layer in the engineering implementation is approximately 614,400 multiplied by 4 bytes, or 2.4 megabytes, which is negligible considering that current embedded controllers are generally equipped with 4 gigabytes of memory. The above engineering indicators show that the system described in this invention has good engineering feasibility on a general-purpose embedded controller and can be deployed without any dedicated hardware.

[0061] This invention describes the scalability of the system. The core data structure of the system is a two-layer cost map, where the basic geometric cost layer and the tracking executability prediction layer are independent of each other, coupled through an integrated cost calculation. This data structure has good scalability and can be easily integrated into existing robot navigation software stacks. When integrated with the navigation stack of the ROS2 operating system, the tracking executability prediction layer can be added as a new cost layer in addition to the existing grid cost layer, injected into the navigation cost map through the existing costmap layer overlay mechanism, without requiring any modification to the existing path search algorithm. When integrated with existing SLAM algorithms, the tracking executability prediction layer can be maintained by the SLAM algorithm in the synchronous update after loop closure detection, and the grid's tracking executability prediction value automatically follows and updates with the SLAM's pose optimization. When integrated with existing motion control software, the tracker only needs to package and send back the pose deviation of each completed grid to the tracking executability update unit; the interface consists of only one function call. In multi-robot collaborative scenarios, the tracking executability prediction layer described in this invention can be synchronized among multiple robots as a shared cost layer. Each robot contributes its own tracking executability evidence, collectively accumulating grid-level tracking difficulty information in the workspace. None of the above four integration methods require any modification to the system itself, demonstrating that the system possesses good scalability and portability.

[0062] A comprehensive analysis of the advantages of the system described in this invention. The advantages of the system described in this invention compared to existing typical routes can be comprehensively analyzed from the following four dimensions. First, in terms of navigation efficiency, the system described in this invention, in the typical workspaces of Examples 1 to 4, shows a reduction in navigation completion time of 26.8%, 17.9%, 17.7%, and 23.2% respectively compared to the existing two-stage serial architecture, demonstrating a significant improvement in navigation efficiency across different workspaces. Second, in terms of tracking accuracy, the system described in this invention, in Examples 1 to 4, shows a reduction in the cumulative average tracking residual of 61.1%, 66.7%, 55.6%, and 62.5% respectively compared to the existing two-stage serial architecture, demonstrating a reduction in tracking residual of over 50% across all typical workspaces, indicating a significant improvement. Third, in terms of engineering deployment difficulty, the system described in this invention does not require the introduction of neural networks, does not require any SysID dynamic model calibration, and does not require architectural modifications to the existing planning-tracking software stack. The newly added interface consists of only one line of function calls, significantly reducing the engineering deployment threshold compared to methods based on neural networks or dynamic model calibration. Fourth, in terms of adaptability, the system described in this invention automatically adapts to changes in grid tracking difficulty within the workspace without any manual adjustment. When temporary obstacles are added to the workspace, it automatically accumulates low tracking executability estimates and avoids them. When existing obstacles are removed from the workspace, the tracking executability estimates naturally recover through a moving average convex combination mechanism. This demonstrates a significant adaptive advantage compared to existing methods that rely on manual annotation or fixed parameters. The above four-dimensional comprehensive analysis shows that the system described in this invention significantly reduces navigation completion time while greatly compressing the cumulative average tracking residual, significantly lowering the engineering deployment threshold, and significantly enhancing adaptive capabilities, achieving a comprehensive improvement over the existing two-stage serial architecture.

[0063] The global optimization mechanism of the system described in this invention is explained. The accumulation process of tracking executability predictions in the system described in this invention can be mathematically viewed as an identification process of workspace dynamics characteristics based on actual execution residuals. The identification results are directly stored in the cost graph in grid-level scalar form. The key difference between this mechanism and existing SysID methods is that the system described in this invention does not require any prior dynamics model assumptions or any additional identification stimulus signals; it directly utilizes tracking residual data from daily navigation tasks for passive identification. The advantage of passive identification is that it does not interrupt the normal workflow, does not occupy additional time, and the identification results accumulate naturally according to the actual usage of the workspace. The key difference between this mechanism and existing reinforcement learning methods is that the system described in this invention does not require any reward function design or any exploration-exploitation tradeoff. The update of the tracking executability prediction depends solely on the actual tracking residuals, the update direction is unique and irreversible, and there is no risk of policy oscillation. The key difference between this mechanism and existing adaptive control methods is that the adaptive granularity of the system described in this invention is grid-level rather than controller parameter-level. Grid-level adaptation can differentiate the adaptation to the dynamic differences in different regions of the workspace, which is more granular and accurate than global adaptation at the controller parameter level. The above comparison shows that the dynamic identification mechanism of the system described in this invention has unique engineering advantages over the existing SysID, reinforcement learning and adaptive control methods.

[0064] The structure of the end-to-end integrated planning and tracking trackless navigation system described in this invention is as follows: Figure 11 As shown. The system of this invention includes five functional units: a cost graph initialization unit, a global planning unit, a tracing execution unit, a tracing executability update unit, and a closed-loop control unit. The cost graph initialization unit executes step S0, the global planning unit executes step S1, the tracing execution unit executes step S2, the tracing executability update unit executes step S3, and the closed-loop control unit executes step S4. The system of this invention can be deployed on a general-purpose embedded controller. A typical embodiment uses an embedded controller based on an ARM Cortex-A series processor, running the ROS2 operating system, and with at least 4 gigabytes of memory. The overall single-round execution time of all five functional units does not exceed 80 milliseconds.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention. The core scope of protection of the system described in the present invention is defined by the claims. The parameter values ​​given in the specific embodiments are merely illustrative examples and are not intended to limit the scope of protection. After reading the specific embodiments of the system described in the present invention, those skilled in the art should be able to make appropriate adjustments to the parameter values ​​and specific implementation methods of the system described in the present invention without creative effort, and apply them to other similar workspaces not explicitly stated in the present invention while retaining the core ideas of the system described in the present invention. The six typical workspaces given in embodiments 1 to 6 of the system described in the present invention are merely examples. The system described in the present invention can be applied to, but is not limited to, fields such as automated guided vehicles, autonomous mobile robots, indoor service robots, welcoming robots, delivery robots, and multi-machine collaborative logistics systems. Any trackless autonomous navigation system involving a two-stage serial architecture of planning and tracking can benefit from the end-to-end integrated feedback mechanism of the system described in the present invention.

Claims

1. A trackless navigation method with end-to-end integrated planning and tracking, characterized in that, The method is operated according to the following steps: S0, obtain the initial cost map of the workspace. The initial cost map includes a basic geometric cost layer and a tracking executability prediction layer. The basic geometric cost layer is a cost distribution obtained according to the obstacle distance field. Each grid in the tracking executability prediction layer independently maintains a tracking executability prediction value. The initial value of the tracking executability prediction value is 1 and falls within [0, 1]. S1, the global planner searches for candidate paths for the current navigation task on the cost map using a preset path search algorithm. For each grid cell traversed by each candidate path, the basic geometric cost is divided by the estimated tracking executability of the grid cell to obtain the integrated cost component. The integrated cost components of all grid cells are summed to obtain the integrated cost of the candidate path. The candidate path with the minimum integrated cost is taken as the optimal path for this round. S2, the tracker performs motion control on the trackless vehicle according to the optimal path of this round using a preset tracking control algorithm. For each completed grid, the trackless vehicle's actual pose and the path reference pose are recorded as the lateral deviation and heading deviation, and the residuals of this grid in this round are synthesized according to the preset residual synthesis function. S3, For each completed grid, the tracking residual of this round is mapped to the instantaneous value of tracking executability according to a preset inverse proportional mapping function, wherein the preset inverse proportional mapping function is a monotonically decreasing function; The updated tracking executability estimate is obtained by combining the tracking executability estimate and the instantaneous tracking executability value using a convex combination based on the preset moving average weight. S4, if the navigation task is not completed, return to S1 to trigger the next closed loop; When the navigation task is completed, the completion time, cumulative trajectory length and cumulative average tracking residual of this navigation are output, and the updated tracking executability prediction layer is retained for subsequent tasks.

2. The end-to-end integrated planning and tracking trackless navigation method according to claim 1, characterized in that, The preset residual synthesis function is a weighted sum of the absolute values ​​of the lateral deviation and the absolute values ​​of the heading deviation. The weight of the lateral deviation is not less than 1.0 and not greater than 5.0, and the weight of the heading deviation is not less than 0.5 and not greater than 2.

0.

3. The end-to-end integrated planning and tracking trackless navigation method according to claim 1, characterized in that, The preset inverse proportional mapping function adopts any of the following: linear inverse proportional mapping, exponential inverse proportional mapping, or normalized inverse proportional mapping based on hyperbolic tangent; The cutoff residual threshold of the linear inverse mapping is no less than 10 cm and no more than 50 cm.

4. The trackless navigation method for end-to-end integrated planning and tracking according to claim 1, characterized in that, The preset moving average weight is not less than 0.05 and not greater than 0.

30. The convex combination is executed as follows: the updated tracking executability estimate is equal to the instantaneous tracking executability value multiplied by the moving average weight plus the tracking executability estimate before the update multiplied by 1 minus the moving average weight.

5. The end-to-end integrated planning and tracking trackless navigation method according to claim 1, characterized in that, The values ​​of the basic geometric cost layer are obtained as follows: the maximum cost value is taken for the grid cells identified as fixed obstacles; For grids whose distance from a fixed obstacle grid is no greater than a preset safety distance threshold, the maximum cost value is multiplied by an attenuation coefficient, and the attenuation coefficient decreases monotonically with the distance from the fixed obstacle grid. For grids that are more than the preset safe distance threshold from a fixed obstacle grid, a zero cost is applied.

6. The end-to-end integrated planning and tracking trackless navigation method according to claim 1, characterized in that, The preset path search algorithm adopts any one of the following: A* search algorithm, Dijkstra search algorithm, fast traversal algorithm, or fast expanding random tree algorithm with random sampling.

7. The end-to-end integrated planning and tracking trackless navigation method according to claim 1, characterized in that, The preset tracking control algorithm adopts any one of the following: pure tracking method, model predictive control method, dynamic window method, or trajectory tracking method based on feedback linearization.

8. The end-to-end integrated planning and tracking trackless navigation method according to claim 1, characterized in that, For graticules whose estimated tracking executability is less than a preset extremely low tracking executability threshold, when calculating the integrated cost component of the graticule in S1, the basic geometric cost is multiplied by a preset temporary masking coefficient. The preset extremely low tracking executability threshold is no greater than 0.10, and the preset temporary masking coefficient is no less than 5.0 and no greater than 20.0, so that the path search algorithm automatically avoids graticules that are historically extremely difficult to track in the next round of planning.

9. The end-to-end integrated planning and tracking trackless navigation method according to claim 1, characterized in that, For grids that have accumulated more than 50 consecutive rounds without execution, the tracking executability estimate of the grid is reset to an initial value of 1.

10. A trackless navigation system with end-to-end integrated planning and tracking, characterized in that... ,include: A two-layer cost map initialization module is used to obtain an initial cost map of the workspace. The initial cost map includes a basic geometric cost layer and a tracking executability prediction layer. The basic geometric cost layer is a cost distribution obtained according to the obstacle distance field. Each grid in the tracking executability prediction layer independently maintains a tracking executability prediction value. The initial value of the tracking executability prediction value is 1 and falls within [0, 1]. The integrated cost planning module is used by the global planner to search for candidate paths for the current navigation task on the cost map using a preset path search algorithm. For each grid cell traversed by each candidate path, the basic geometric cost is divided by the estimated tracking executability of the grid cell to obtain the integrated cost component. The integrated cost components of all grid cells are summed to obtain the integrated cost of the candidate path. The candidate path with the minimum integrated cost is selected as the optimal path for this round. The tracking execution and deviation recording module is used by the tracker to perform motion control on the trackless vehicle according to the optimal path of the current round and a preset tracking control algorithm. For each completed grid, it records the lateral deviation and heading deviation between the actual pose of the trackless vehicle and the path reference pose, and synthesizes them into the tracking residual of the current grid for the current round according to a preset residual synthesis function. The tracking executability estimate update module is used to map the tracking residual of the current round to the tracking executability instantaneous value for each completed grid according to a preset inverse proportional mapping function. The preset inverse proportional mapping function is a monotonically decreasing function. The updated tracking executability estimate is obtained by convexly combining the tracking executability estimate and the tracking executability instantaneous value according to a preset moving average weight. The closed-loop control and navigation completion determination module is used to drive the trackless vehicle back to the integrated cost planning module to trigger the next round of closed loop when the navigation task is not completed. When the navigation task is completed, it outputs the completion time of this navigation, the cumulative trajectory length and the cumulative average tracking residual, and retains the updated tracking executability prediction layer for subsequent tasks.

Images