Training evolution method for real-time mapping and positioning capability of unmanned system

Abstract

The invention discloses a training evolution method for real-time mapping and positioning capability of an unmanned system, and relates to the field of unmanned system simulation training and positioning mapping, and the method comprises the steps: constructing a city-level simulation scene set covering a plurality of city environment characteristics, and carrying out the parameterization and randomization configuration; determining a hyper-parameter group of a definition algorithm and a constraint domain thereof; driving the unmanned system to execute multiple rounds of repeated simulation tests according to a preset task script, and collecting related data; forming a multi-index evaluation vector, and further generating a comprehensive score; iteratively updating the hyper-parameter group to obtain a candidate hyper-parameter group, and writing information corresponding to the current iteration round into a capability evolution knowledge base; and searching the hyper-parameter group from the capability evolution knowledge base as an initialization parameter to carry out parameter migration, carrying out continuous iterative optimization on the candidate hyper-parameter group by adopting a curriculum type difficulty propulsion mechanism until a preset convergence condition and a real-time constraint are met, and outputting an optimal hyper-parameter group and an applicable scene domain thereof.

Description

Technical Field

[0001] This invention relates to the field of unmanned system simulation training and localization mapping technology, and in particular to a training evolution method for real-time mapping and localization capabilities of unmanned systems. Background Technology

[0002] When unmanned systems perform tasks such as inspection, delivery, and security in urban environments, they rely on real-time mapping and localization algorithms to achieve autonomous positioning and environmental modeling. Urban environments are typically characterized by complex road structures, frequent building occlusion, high repetition of textures and geometric features, significant changes in lighting and weather, and dense dynamic traffic participants. These factors all place high demands on the robustness and adaptability of real-time mapping and localization algorithms.

[0003] In existing technologies, parameter tuning for real-time mapping and positioning systems largely relies on expert experience, achieved through repeated trials and manual adjustments in limited scenarios. This method has the following limitations: First, the parameter tuning process is usually based on a single or a few typical scenarios, which can easily lead to parameter settings being overly adapted to specific environments and lacking generalization ability across scenarios and multiple working conditions.

[0004] Secondly, real-time mapping and positioning systems must also meet strict real-time performance and computational resource constraints. However, current parameter tuning methods often focus on accuracy optimization and tend to ignore the balance between real-time performance and resource consumption, making it difficult for the algorithm to meet performance requirements in actual deployment.

[0005] Furthermore, existing methods lack a systematic recording and reuse mechanism for the "scenario-parameter-performance" relationship, resulting in the need to conduct a large number of parameter tuning experiments repeatedly in different projects or new scenarios, leading to high tuning costs, long cycles, and low efficiency in knowledge accumulation and transfer.

[0006] Therefore, how to provide a systematic and automated method to optimize and evolve parameters across different scenarios, and improve the adaptability and stability of unmanned system mapping and positioning algorithms in different urban environments while ensuring real-time performance and resource constraints, has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a training and evolution method for real-time mapping and localization capabilities of unmanned systems, comprising the following steps: Construct a set of city-level simulation scenarios covering various urban environmental characteristics, and perform parameterization and randomization configuration of scenario elements; Determine the real-time mapping and localization algorithm to be trained and evolved, define the hyperparameter set and its constraint domain of the algorithm, and establish the feasibility verification rules for the hyperparameter set. In the set of city-level simulation scenarios, the unmanned system is driven to perform multiple rounds of repeated simulation experiments according to the preset task script, and trajectory data, map data and runtime resource data are collected. Based on the collected data, multiple evaluation indicators are calculated to form a multi-indicator evaluation vector, and a comprehensive score is generated based on the multi-indicator evaluation vector. By combining error-pattern-based automatic parameter tuning rules and multi-objective optimization strategies, the hyperparameter group is iteratively updated to obtain candidate hyperparameter groups, and the scenario configuration information, hyperparameter group and performance indicators corresponding to the current iteration round are written into the capability evolution knowledge base. During cross-scenario training, hyperparameter sets are retrieved from the capability evolution knowledge base based on scenario similarity as initialization parameters for parameter migration. A course-based difficulty advancement mechanism is used to continuously iterate and optimize the candidate hyperparameter sets until the preset convergence conditions and real-time constraints are met, and the optimal hyperparameter set and its applicable scenario domain are output.

[0008] Furthermore, the step of constructing a city-level simulation scene set covering multiple urban environmental characteristics includes: Construct basic topology models for multiple city simulation scenarios, wherein the basic topology models include at least road structure, intersection type, building cluster form and landmark distribution; Lighting and weather conditions can be generated in a controllable manner, and the density, speed distribution, and behavior patterns of dynamic traffic participants can be parameterized. The sensor degradation factors are parameterized, including noise intensity, external parameter disturbance, time synchronization error, frame loss probability, or occlusion probability. A random parameter set is generated based on a random seed generator, and a level parameter set is generated according to a preset challenge level to form a set of simulation scenarios containing multiple cities, multiple difficulty levels, and multiple working conditions.

[0009] Furthermore: The random parameter set of the lighting and weather conditions includes cloud type, cloud height, sun azimuth or wind direction and speed. The group of level parameters for the lighting and weather conditions includes cloud cover, visibility, intensity of rain, snow, fog, or haze, or intensity of road surface reflectivity. Specifically, the parameters in the level parameter group are normalized and divided into multiple challenge levels for use by the course-based difficulty progression mechanism.

[0010] Furthermore, the hyperparameter group includes at least front-end odometry hyperparameters, keyframe or keypoint management hyperparameters, loop closure detection hyperparameters, back-end graph optimization hyperparameters, and map update hyperparameters.

[0011] Furthermore: The front-end odometer hyperparameters include feature extraction threshold, matching window size, voxel resolution, or motion model weights. The backend graph optimization hyperparameters include sliding window size, robust kernel parameter, edge weight ratio, upper limit of iterations, or convergence threshold.

[0012] Furthermore, the feasibility verification rules include: The real-time constraint is used to limit the single-frame processing delay or real-time factor from not exceeding a preset threshold. Resource usage constraints are used to limit CPU and GPU utilization, memory usage, or bandwidth usage to no more than preset thresholds; Stability constraints are used to determine infeasible parameter sets for trajectory divergence, graph optimization non-convergence, or abnormal map structure.

[0013] Furthermore, the preset task script includes at least: fixed-route cruise task, random exploration task, occlusion crossing task, dynamic interference task, and sensor degradation condition task; In this process, multiple repeated trials are conducted for each task under different random seeds to obtain statistically stable evaluation results.

[0014] Furthermore, the multiple evaluation indicators include at least: absolute trajectory error, relative pose error, drift rate, loop closure consistency error, map integrity, map redundancy, single-frame processing latency, real-time factor, and resource consumption indicators.

[0015] Furthermore, the step of generating the comprehensive score includes: The weighted terms are obtained by weighting and summing the various evaluation indicators. Penalties are imposed based on violations of the real-time constraints and resource consumption constraints; The comprehensive score is composed of the weighted terms and the penalty terms.

[0016] Furthermore, the automatic parameter tuning rules based on error patterns include: When an error pattern of trajectory drift or increased cumulative error is detected, parameter tuning actions are performed, such as increasing the loop closure trigger strength, increasing the back-end constraint weight, or adjusting the relocation threshold. When error patterns such as map jitter, repeated mapping, or increased closed-loop conflicts are identified, parameter tuning actions such as tightening keyframe insertion conditions, increasing robust kernel strength, or reducing noise-sensitive parameters are performed. When an error pattern is identified where the single-frame processing latency exceeds a threshold, parameter tuning actions are performed, such as reducing the upper limit of the number of features or point clouds, reducing the loop closure detection frequency, or shrinking the back-end sliding window.

[0017] Furthermore, the multi-objective optimization strategy is Bayesian optimization, evolutionary strategy, or hybrid optimization strategy.

[0018] Furthermore, the hybrid optimization strategy includes: First, initial values ​​for candidate hyperparameter sets are generated using the automatic parameter tuning rules based on error patterns, and then a black-box search is performed in the neighborhood of the initial values.

[0019] Furthermore, the specific mechanism for advancing the difficulty of the course is as follows: First, train in low-challenge-level scenes with low dynamics, low occlusion, and good lighting to obtain the basic hyperparameter set; Then, high-challenge scenarios involving highly dynamic traffic participants, strong obstruction, and complex weather are introduced step by step to iteratively optimize the basic hyperparameter set.

[0020] Furthermore, the convergence conditions include the comprehensive score improving by less than a threshold in N consecutive iterations, or the key accuracy indicators reaching the target and the real-time constraints being met, or reaching the maximum number of iterations; and outputting the optimal hyperparameter set and its applicable scenario domain.

[0021] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention establishes a systematic and automated training and evolution framework by constructing a parameterized multi-scenario simulation set, defining hyperparameter constraints and feasibility verification rules, and automatically iterating and tuning parameters based on error patterns and multi-objective optimization. This reduces reliance on human experience and the cost of parameter tuning, and achieves cross-scenario parameter transfer and knowledge accumulation through a capability evolution knowledge base, thereby improving the generalization ability and stability of unmanned system mapping and positioning algorithms in different urban environments. At the same time, the strict coupling of real-time performance and resource constraint verification during the optimization process ensures the feasibility of the obtained hyperparameter set in actual deployment, achieving coordinated optimization of algorithm performance and engineering implementation. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the training evolution method disclosed in this invention; Figure 2 This is a flowchart illustrating the process of constructing a multi-city simulation scene set as disclosed in this invention. Figure 3 This is a schematic diagram of the automatic parameter tuning rules based on error patterns disclosed in this invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] This invention aims to provide a training and evolution method for real-time mapping and localization capabilities of unmanned systems. It can systematically and automatically achieve cross-scenario parameter optimization and evolution, thereby improving the adaptability and stability of unmanned system mapping and localization algorithms in different urban environments while ensuring real-time performance and resource constraints.

[0026] Please see Figure 1 The method mainly includes the following steps: S1. Construct a set of city-level simulation scenarios covering various urban environmental characteristics, and configure the scenario elements parametrically and randomly.

[0027] Those skilled in the art will understand that this step aims to establish a diverse, quantifiable, and controllable simulation testing environment to simulate various challenges that may be encountered in real cities. In specific implementation, it can be further refined into sub-steps S11 to S14 to form a controllable, reproducible, and hierarchical set of city-level simulation scenarios, providing a unified scenario parameter space and indexing method for subsequent cross-scenario training and evolution. The process can be found in [reference needed]. Figure 2 .

[0028] S11. Construct basic topology models for multiple city simulation scenarios.

[0029] Those skilled in the art will understand that multiple representative urban digital twin models (i.e., basic topology models) are constructed in the simulation platform as the basic framework. The basic topology model includes at least road structure, intersection types, building cluster morphology, and landmark distribution. These features are encoded as vectors for subsequent scene similarity calculation and retrieval. Specifically, the basic topology model can be organized using a graph structure, where nodes represent road endpoints, intersections, ramp connections, etc., and edges represent road segments, assigned attributes such as road grade, number of lanes, speed limit, slope, and curvature. Intersection types can be categorized as crosses, T-junctions, roundabouts, grade-separated intersections, and complex multi-branch intersections. Building cluster morphology includes at least statistical features such as building footprint outline, building height distribution, block size, and building density. Landmark distribution includes at least a stable set of feature points that can be used for visual / laser loop retrieval, such as bridges, tunnels, plazas, prominent high-rise buildings, or special structures. These are used for subsequent scene similarity retrieval and parameter migration initialization.

[0030] S12. Generate controllable lighting and weather conditions and parameterize dynamic traffic participants.

[0031] Those skilled in the art will understand that, on a given topology, environmental and dynamic elements are parametrically injected to generate multiple working condition variants under the same urban structure.

[0032] The illumination and weather conditions are generated using a parameter-driven approach. During implementation, illumination parameters include solar azimuth / altitude angle, illumination intensity, and cloud type / altitude; weather parameters include the type and intensity of rain, snow, fog, and haze, as well as wind direction and speed. These parameters can further derive physical quantities affecting sensor observations, such as visibility and road surface slipperiness / reflectivity. To construct a systematic difficulty hierarchy, key parameters such as cloud cover, visibility, and rain / snow intensity can be normalized and divided into several clearly defined challenge levels (e.g., L1 to L5).

[0033] Dynamic traffic participants involve modeling motor vehicles, non-motor vehicles, and pedestrians. During implementation, their parameter settings include their density in the road network (e.g., number of vehicles per kilometer), speed distribution (mean and variance), and behavioral patterns (e.g., following, lane changing, and obstacle avoidance strategies). By adjusting density and behavioral complexity, traffic conditions with different challenge levels—"low dynamic," "medium dynamic," and "high dynamic"—can be created.

[0034] S13. Set sensor degradation elements and form a set of degradation condition parameters.

[0035] Those skilled in the art will understand that this means controllingly injecting degradation factors into the non-ideal state of the simulated sensor in actual deployment. Degradation factors include noise intensity, external parameter disturbances, time synchronization errors, frame loss probability, or occlusion probability, used to construct a controllable set of degradation conditions.

[0036] Specifically, noise intensity can include pixel noise / exposure perturbation / motion blur parameters for camera observations, distance noise / angle noise / echo missing probability for LiDAR observations, and zero bias and random walk parameters for IMUs. External parameter perturbations can be modeled using small rotation and translation perturbations of the relative pose of the sensors, with the perturbation range limited to meet feasibility constraints. Time synchronization errors can be simulated using a fixed bias or time-drift delay model to model asynchronous sampling by multiple sensors. Frame drop probability can be simulated using a Bernoulli process or a segmented frame drop strategy to model communication jitter or computational congestion. Occlusion probability can be achieved through dynamic volume occlusion enhancement or local viewpoint occlusion models to amplify typical degradation scenarios such as urban canyons and traffic occlusion.

[0037] It should be noted that the set of degradation condition parameters formed by S13 should be composable with the environmental and dynamic volume parameters of S12, so that different degradation intensities and different dynamic traffic intensities can be superimposed on the same urban topology to form a more comprehensive "condition Cartesian product".

[0038] S14. Generate random parameter sets and level parameter sets to form the final scene set.

[0039] Those skilled in the art will understand that the control parameters of the above-mentioned elements are divided into two categories: The first type, random parameter group, is controlled by a seed random generator and is mainly used to generate subtle perturbations (including cloud details, initial distribution of local traffic flow, randomness of individual behavior, and occurrence time of local occlusion events) under the same challenge level, in order to improve the statistical stability and de-randomness of the evaluation and avoid overfitting to a specific random state.

[0040] The second category, the level parameter group, is used to control the level of scene challenge. It can include at least macro-level difficulty factors that significantly affect real-time mapping and positioning performance, such as building occlusion intensity / street narrowness, dynamic traffic density, visibility / rain, snow, fog and haze intensity, road surface reflectivity, and sensor degradation intensity.

[0041] Finally, by selecting city topology and challenge level, and sampling within the parameter range of that level using random seeds, a large-scale simulation scenario set containing multiple cities, multiple difficulty levels, and multiple working conditions is generated, with each scenario accompanied by a complete structured configuration vector.

[0042] S2. Determine the real-time mapping and localization algorithm to be trained and evolved, define the hyperparameter set and its constraint domain of the algorithm, and establish the feasibility verification rules for the hyperparameter set.

[0043] During implementation, a real-time mapping and localization algorithm framework to be optimized is selected, such as a SLAM framework based on laser-vision-inertial navigation fusion. A set of hyperparameters to be optimized is defined, which should at least cover front-end odometry hyperparameters, keyframe or keypoint management hyperparameters, loop closure detection hyperparameters, back-end graph optimization hyperparameters, and map update hyperparameters. Specifically, front-end odometry hyperparameters include feature extraction threshold, matching window size, voxel filtering resolution, and motion model weights; keyframe or keypoint management hyperparameters include insertion threshold, minimum interval, and redundancy removal conditions; loop closure detection hyperparameters include trigger threshold, candidate number, and geometric verification threshold; back-end graph optimization hyperparameters include sliding window size, robust kernel function parameters, weight ratios of various constraint edges, upper limit of iterations, and convergence threshold; and map update hyperparameters include update frequency and fusion strategy parameters. Simultaneously, reasonable value ranges, i.e., constraint domains, are defined for these hyperparameters.

[0044] In a further proposed solution, to prevent the optimization process from generating parameters that are impractical for deployment, feasibility verification rules are established. These rules primarily include real-time constraints, resource consumption constraints, and stability constraints.

[0045] Real-time constraints require that the processing latency or average processing time (real-time factor) of a single frame must be lower than a preset threshold (e.g., 100ms) to limit the processing latency or real-time factor of a single frame from exceeding the preset threshold.

[0046] Resource usage constraints are used to limit CPU and GPU utilization, memory usage, or bandwidth usage to no more than preset thresholds; Stability constraints monitor the simulation process to determine whether anomalies occur, such as trajectory divergence, failure of backend optimization to converge, or serious structural errors in the generated map. Parameter sets that violate any constraint will be marked as "infeasible." In other words, stability constraints are used to determine infeasible parameter sets due to trajectory divergence, failure of graph optimization to converge, or abnormal map structure.

[0047] S3. In a city-level simulation scenario set, drive the unmanned system to perform multiple rounds of repeated simulation tests according to the preset task script, and collect trajectory data, map data and runtime resource data.

[0048] Those skilled in the art will understand that, to comprehensively evaluate performance, multiple preset task scripts are designed, such as: cruising along a fixed route, randomly exploring within a region, traversing areas obstructed by buildings / vehicle traffic, driving through dense dynamic interference, and operating under sensor degradation conditions. For each set of hyperparameters, the unmanned system model is driven to perform these tasks in a selected scenario. To ensure the reliability of the results, each task-scenario combination is repeated multiple times under different random seeds. During the experiments, trajectory data, map data, runtime resource data, and process logs are collected simultaneously.

[0049] The trajectory data consists of the trajectory estimated by the algorithm and the ground truth trajectory provided by the simulation platform; the map data consists of point cloud maps, raster maps, or semantic maps constructed by the algorithm in real time; the runtime resource data includes single-frame processing time, CPU / GPU utilization, memory usage, etc.; and the process log includes loop closure trigger / removal events, tracking loss counts, relocation events, backend optimization residuals, etc.

[0050] S4. Calculate multiple evaluation indicators based on the collected data to form a multi-indicator evaluation vector, and generate a comprehensive score based on the multi-indicator evaluation vector.

[0051] Those skilled in the art will understand that S4 involves offline analysis of the data collected in S3, calculating a comprehensive set of evaluation metrics to form an evaluation vector. In implementation, these evaluation metrics include at least absolute trajectory error, relative pose error, drift rate, loop closure consistency error, map integrity, map redundancy, single-frame processing latency, real-time factor, and resource consumption metrics.

[0052] Subsequently, a comprehensive score is generated. Specifically, each indicator is first normalized, then appropriate weights are assigned based on task importance, and a weighted sum is performed to obtain the base score. Afterward, a constraint penalty is introduced: if the real-time constraints or resource consumption constraints defined in S2 are violated, the corresponding penalty points are deducted from the base score, or the comprehensive score for that experiment is directly set to unsatisfactory. This ensures that the optimization process always focuses on deployable solutions.

[0053] S5. Combining error-mode-based automatic parameter tuning rules and multi-objective optimization strategies, the hyperparameter group is iteratively updated to obtain candidate hyperparameter groups. The scenario configuration information, hyperparameter group and performance indicators corresponding to the current iteration round are written into the capability evolution knowledge base.

[0054] Optionally, the multi-objective optimization strategy can be Bayesian optimization, evolutionary optimization, or a hybrid optimization strategy. The hybrid optimization strategy includes: first generating initial values ​​for candidate hyperparameter sets using automatic hyperparameter tuning rules based on error patterns, and then performing a black-box search within the neighborhood of these initial values.

[0055] Those skilled in the art will understand that S5 is the core of automatic parameter tuning, and its logic can be found in [reference needed]. Figure 3 The diagram shows an automatic parameter tuning rule based on error patterns.

[0056] S51. Error Pattern Recognition. Analyze the multi-index evaluation vector obtained in S4 and the operation logs collected in S3. By setting thresholds or pattern matching rules, performance defects are classified into specific error patterns, such as: trajectory drift (continuous increase in ATE / RPE), closed-loop conflict (errors causing map distortion), map jitter / redundancy (unstable map or many duplicate elements), and real-time performance exceeding limits (processing latency exceeds the standard).

[0057] S52-S55, Generate targeted parameter tuning actions. Based on the identified error pattern, trigger a predefined parameter tuning rule library to generate a set of specific parameter adjustment directions. During implementation, the automatic parameter tuning rules based on the error pattern specifically include: For trajectory drift type: it is recommended to increase the sensitivity of loop closure detection (e.g., lower the trigger threshold), increase the weight of loop closure constraints in backend optimization, or adjust the relocation strategy parameters.

[0058] For closed-loop conflict types: it is recommended to tighten the geometric verification conditions for loop closures, raise the threshold for screening loop closure candidates, or enhance the strength of the robust kernel function in the back-end optimization to suppress erroneous loop closures.

[0059] For map jitter / redundancy issues: it is recommended to tighten keyframe insertion conditions (such as increasing the parallax change threshold), increase the incentive for keyframe removal, or reduce the confidence level of noise-sensitive feature parameters.

[0060] For real-time performance exceeding limits: it is recommended to reduce the number of features processed by the front end, reduce the frequency of loop closure detection, or reduce the size of the sliding window for back-end optimization.

[0061] S56. Determine the update magnitude and perform constraint projection. Based on the severity of the error or the decrease in the overall score, determine the adjustment step size for the above parameter tuning actions (fixed step size, adaptive step size, etc., can be used). Apply the adjustment magnitude to the current hyperparameter set to obtain a new set of parameter values. Then, immediately perform constraint projection on this new set of parameters, that is, check whether it is still within the constraint domain defined in step S2, and pre-check whether it may violate the feasibility verification rules. If it does not meet the requirements, truncate or roll back to the previous feasible parameter.

[0062] S57. Generate and screen candidate hyperparameter groups. Typically, a single iteration generates multiple candidate hyperparameter groups. For example, the main rule adjustment group and its neighborhood perturbation group. Perform a rapid feasibility check on all candidate groups (based on a simplified model or historical data), eliminating those that are clearly infeasible. The remaining feasible candidate groups will proceed to the next round of simulation experiments (steps S3-S4) for a complete evaluation.

[0063] In a further proposed approach, after each complete iteration (from parameter setting to performance evaluation), a complete record entry will be generated, including: scenario configuration vector, hyperparameter group used, task script identifier, multi-metric evaluation vector, comprehensive score, and key operation log summary. This entry will be written into the capability evolution knowledge base, forming searchable and analyzable historical experience data.

[0064] The rule-based hyperparameter tuning described above can be combined with black-box optimization algorithms such as Bayesian optimization and evolutionary strategies to form a hybrid strategy. In one specific embodiment, rules are first used to quickly locate the direction of hyperparameter tuning, and then Bayesian optimization is used to perform a fine search within the parameter neighborhood suggested by the rules, so as to balance the intuitiveness of hyperparameter tuning and search efficiency.

[0065] S6. During cross-scenario training, hyperparameter sets are retrieved from the capability evolution knowledge base based on scenario similarity as initialization parameters for parameter transfer. A course-style difficulty advancement mechanism is adopted to continuously iterate and optimize the candidate hyperparameter sets until the preset convergence conditions and real-time constraints are met, and the optimal hyperparameter set and its applicable scenario domain are output.

[0066] Those skilled in the art will understand that step S6 aims to achieve cross-scenario generalization and progressive improvement of capabilities. Specifically, when faced with a completely new urban scene, the similarity between its scene configuration vector and existing scenes in the knowledge base is first calculated. From the scene records with the highest similarity, the hyperparameter set with the highest comprehensive score is retrieved and used as the initialization parameters for training the new scene. This avoids tuning parameters from scratch and greatly improves efficiency.

[0067] In this scheme, the course-based difficulty progression mechanism works as follows: First, training is conducted in low-challenge scenarios with low dynamics, low occlusion, and good lighting, allowing the hyperparameter set to quickly converge to a version with good basic performance, thus obtaining the basic hyperparameter set; then, the scenario difficulty is gradually increased, such as gradually introducing high-challenge scenarios with high dynamic traffic participants, strong occlusion, and complex weather, and even adding sensor degradation. At each level of difficulty, iterative optimization is performed using the optimal parameters obtained in the previous level as the starting point, so that the algorithm gradually adapts to more demanding conditions.

[0068] In a further proposed approach, the iterative process continues until predefined convergence conditions are met. These conditions include: the overall score improving by less than a threshold over N consecutive iterations; key accuracy indicators (such as ATE) reaching target values ​​while real-time constraints are satisfied; or reaching the maximum number of iterations. Finally, one or more optimal hyperparameter sets are output, with their validated applicable scenario domains clearly labeled. For example, suitable for daytime scenarios in urban areas with moderate-density dynamic traffic. This enables the continuous and traceable evolution of the unmanned system's mapping and positioning capabilities as the scenario library expands.

[0069] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A training and evolution method for real-time mapping and localization capabilities of unmanned systems, characterized in that, Includes the following steps: Construct a set of city-level simulation scenarios covering various urban environmental characteristics, and perform parameterization and randomization configuration of scenario elements; Determine the real-time mapping and localization algorithm to be trained and evolved, define the hyperparameter set and its constraint domain of the algorithm, and establish the feasibility verification rules for the hyperparameter set. In the set of city-level simulation scenarios, the unmanned system is driven to perform multiple rounds of repeated simulation experiments according to the preset task script, and trajectory data, map data and runtime resource data are collected. Based on the collected data, multiple evaluation indicators are calculated to form a multi-indicator evaluation vector, and a comprehensive score is generated based on the multi-indicator evaluation vector. By combining error-pattern-based automatic parameter tuning rules and multi-objective optimization strategies, the hyperparameter group is iteratively updated to obtain candidate hyperparameter groups, and the scenario configuration information, hyperparameter group and performance indicators corresponding to the current iteration round are written into the capability evolution knowledge base. During cross-scenario training, hyperparameter sets are retrieved from the capability evolution knowledge base based on scenario similarity as initialization parameters for parameter migration. A course-based difficulty advancement mechanism is used to continuously iterate and optimize the candidate hyperparameter sets until the preset convergence conditions and real-time constraints are met, and the optimal hyperparameter set and its applicable scenario domain are output.

2. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The steps for constructing a city-level simulation scene set covering multiple urban environmental characteristics include: Construct basic topology models for multiple city simulation scenarios, wherein the basic topology models include at least road structure, intersection type, building cluster form and landmark distribution; Lighting and weather conditions can be generated in a controllable manner, and the density, speed distribution, and behavior patterns of dynamic traffic participants can be parameterized. The sensor degradation factors are parameterized, including noise intensity, external parameter disturbance, time synchronization error, frame loss probability, or occlusion probability. A random parameter set is generated based on a random seed generator, and a level parameter set is generated according to a preset challenge level to form a set of simulation scenarios containing multiple cities, multiple difficulty levels, and multiple working conditions.

3. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 2, characterized in that: The random parameter set of the lighting and weather conditions includes cloud type, cloud height, sun azimuth or wind direction and speed. The group of level parameters for the lighting and weather conditions includes cloud cover, visibility, intensity of rain, snow, fog, or haze, or intensity of road surface reflectivity. Specifically, the parameters in the level parameter group are normalized and divided into multiple challenge levels for use by the course-based difficulty progression mechanism.

4. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The hyperparameter group includes at least front-end odometer hyperparameters, keyframe or keypoint management hyperparameters, loop closure detection hyperparameters, back-end graph optimization hyperparameters, and map update hyperparameters.

5. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 4, characterized in that: The front-end odometer hyperparameters include feature extraction threshold, matching window size, voxel resolution, or motion model weights. The backend graph optimization hyperparameters include sliding window size, robust kernel parameter, edge weight ratio, upper limit of iterations, or convergence threshold.

6. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The feasibility verification rules include: The real-time constraint is used to limit the single-frame processing delay or real-time factor from not exceeding a preset threshold. Resource usage constraints are used to limit CPU and GPU utilization, memory usage, or bandwidth usage to no more than preset thresholds; Stability constraints are used to determine infeasible parameter sets for trajectory divergence, graph optimization non-convergence, or abnormal map structure.

7. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The preset task script includes at least: fixed-route cruise task, random exploration task, occlusion crossing task, dynamic interference task, and sensor degradation condition task. In this process, multiple repeated trials are conducted for each task under different random seeds to obtain statistically stable evaluation results.

8. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 6, characterized in that, The evaluation metrics include at least: absolute trajectory error, relative pose error, drift rate, loop closure consistency error, map integrity, map redundancy, single-frame processing latency, real-time factor, and resource usage metrics.

9. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 8, characterized in that, The steps for generating the comprehensive score include: The weighted terms are obtained by weighting and summing the various evaluation indicators. Penalties are imposed based on violations of the real-time constraints and resource consumption constraints; The comprehensive score is composed of the weighted terms and the penalty terms.

10. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The automatic parameter tuning rules based on error patterns include: When an error pattern of trajectory drift or increased cumulative error is detected, parameter tuning actions are performed, such as increasing the loop closure trigger strength, increasing the back-end constraint weight, or adjusting the relocation threshold. When error patterns such as map jitter, repeated mapping, or increased closed-loop conflicts are identified, parameter tuning actions such as tightening keyframe insertion conditions, increasing robust kernel strength, or reducing noise-sensitive parameters are performed. When an error pattern is identified where the single-frame processing latency exceeds a threshold, parameter tuning actions are performed, such as reducing the upper limit of the number of features or point clouds, reducing the loop closure detection frequency, or shrinking the back-end sliding window.

11. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The multi-objective optimization strategy is Bayesian optimization, evolutionary strategy, or hybrid optimization strategy.

12. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 11, characterized in that, The hybrid optimization strategy includes: First, initial values ​​for candidate hyperparameter sets are generated using the automatic parameter tuning rules based on error patterns, and then a black-box search is performed in the neighborhood of the initial values.

13. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The specific mechanism for advancing difficulty in the curriculum is as follows: First, train in low-challenge-level scenes with low dynamics, low occlusion, and good lighting to obtain the basic hyperparameter set; Then, high-challenge scenarios involving highly dynamic traffic participants, strong obstruction, and complex weather are introduced step by step to iteratively optimize the basic hyperparameter set.

14. The training and evolution method for real-time mapping and localization capabilities of unmanned systems according to claim 1, characterized in that, The convergence conditions include the comprehensive score improving by less than a threshold in N consecutive iterations, or the key accuracy indicators reaching the target and the real-time constraints being met, or reaching the maximum number of iterations; and outputting the optimal hyperparameter set and its applicable scenario domain.

Images