Anti-skid control method and device for AGV complex road fixed route, computer device and storage medium

Abstract

The application discloses a method and device for anti-skid control of AGV complex road fixed route, computer equipment and storage medium, the method comprises the following steps: constructing a fixed route tracking environment model and a factory floor environment model, both of which constitute an offline training environment model; training a fixed route tracking strategy library and a skid compensation strategy library; obtaining the task information of AGV executing the fixed route, and real-time collecting the road surface image or texture feature of the position where the AGV is located, the motion state information of the AGV itself and the load information; matching the corresponding initial skid compensation strategy; online fine-tuning the initial skid compensation strategy to generate a dynamic skid compensation strategy; generating a target AGV control instruction to form a closed-loop control until the AGV completes the tracking task of the whole fixed route. Through the implementation of the method of the application, the control accuracy, running stability and environmental adaptability of AGV in the real industrial scene when executing the fixed route task can be improved.

Description

Technical Field

[0001] This invention relates to the field of AGV anti-slip control technology, and more specifically to anti-slip control methods, devices, computer equipment, and storage media for AGVs on fixed routes in complex road surfaces. Background Technology

[0002] In the field of motion control technology for Automated Guided Vehicles (AGVs), it is particularly suitable for typical application scenarios with extremely high requirements for operational accuracy and safety, such as factory automated production lines, intelligent warehousing and logistics, port container loading and unloading, and intelligent manufacturing.

[0003] In the scenarios described above, AGVs typically need to perform high-frequency, high-reliability material handling or transportation tasks along pre-set fixed routes. However, the ground conditions in actual operating environments are often complex and variable, such as the presence of oil stains, water accumulation, metal shavings, ramps, or areas where different materials are spliced ​​together, causing sudden changes in the road surface adhesion coefficient. At the same time, the weight and distribution of the goods carried by the AGV may also change dynamically. These factors combined can easily lead to problems such as drive wheel slippage, vehicle instability, and trajectory deviation.

[0004] Currently, the mainstream AGV path tracking control methods in industry mainly rely on classical proportional-integral-derivative (PID) control or model predictive control (MPC). Although these methods perform well on ideal or structured road surfaces, their control performance deteriorates significantly under unstructured, low-adhesion, or dynamically changing complex road surfaces. Specifically, PID controllers lack the ability to adapt to system nonlinearity and external disturbances; while MPC, although possessing some feedforward optimization capabilities, relies on accurate dynamic models, making it difficult to accurately model in real time when road friction characteristics are unknown or abrupt, resulting in control command lag or overshoot, and failing to effectively suppress slippage.

[0005] More importantly, existing technologies generally treat path tracking and anti-slip control as a single objective, failing to achieve decoupling and coordinated regulation of path tracking and anti-slip control at the strategy level. When slippage occurs, the system often can only respond passively, making it difficult to proactively prevent or compensate for slippage risks while ensuring trajectory tracking accuracy. This limits the large-scale deployment of AGVs in high-precision positioning (e.g., within ±2cm) and high-safety-level scenarios.

[0006] Therefore, there is an urgent need for a new control method that can balance high-precision tracking of fixed routes with active anti-skid on complex road surfaces, in order to overcome the bottlenecks of existing technologies in terms of dynamic working condition adaptability, control robustness and real-time response capability, and provide key technical support for the stable, efficient and safe operation of AGV clusters in real industrial environments. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide an anti-skid control method, device, computer equipment and storage medium for AGV fixed routes on complex road surfaces.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] Anti-skid control methods for fixed routes on complex road surfaces used by AGVs include:

[0010] S110 constructs a fixed route tracking environment model based on a preset fixed running route, and constructs a factory floor environment model based on actual measured data of slippery road surfaces in the actual factory scenario. The two together constitute an offline training environment model.

[0011] S120, in the offline training environment model, uses deep reinforcement learning algorithm to train and generate a fixed route tracking strategy library and a slippage compensation strategy library respectively.

[0012] S130: Obtain task information of AGV executing a fixed route, and collect in real time the road surface image or texture features of the current AGV location, AGV's own motion status information, and load information;

[0013] S140 identifies the current road surface type based on road surface images or texture features, and combines the AGV's own motion state information and load information to match the corresponding initial slippage compensation strategy from the slippage compensation strategy library;

[0014] S150, based on the real-time acquisition of the AGV's actual motion response and road surface image or texture features, fine-tunes the initial slippage compensation strategy online to generate a dynamic slippage compensation strategy that adapts to the current working conditions.

[0015] S160: The basic control instructions output by the fixed route tracking strategy library based on the current path point are merged with the compensation control instructions output by the dynamic slippage compensation strategy to generate the target AGV control instructions.

[0016] S170, repeat steps S130 to S160 to form a closed-loop control until the AGV completes the tracking task of the entire fixed route.

[0017] The present invention also provides an anti-skid control device for fixed routes on complex road surfaces of AGVs, comprising:

[0018] The building unit is used to build a fixed route tracking environment model based on a preset fixed running route, and to build a factory floor environment model based on measured data of slippery road surfaces in actual factory scenarios. Together, the two constitute an offline training environment model.

[0019] The generation unit is used to train and generate a fixed route tracking strategy library and a slippage compensation strategy library respectively in the offline training environment model using deep reinforcement learning algorithms.

[0020] The acquisition and collection unit is used to acquire task information of the AGV executing a fixed route, and to collect in real time the road surface image or texture features of the current AGV location, the AGV's own motion status information, and load information;

[0021] The identification and matching unit is used to identify the current road surface type based on the road surface image or texture features, and to match the corresponding initial slippage compensation strategy from the slippage compensation strategy library by combining the AGV's own motion state information and load information.

[0022] The acquisition and fine-tuning unit is used to fine-tune the initial slippage compensation strategy online based on the real-time acquired AGV actual motion response and road surface image or texture features, and generate a dynamic slippage compensation strategy that adapts to the current working conditions.

[0023] The fusion unit is used to fuse the basic control instructions output by the fixed route tracking strategy library based on the current path point with the compensation control instructions output by the dynamic slippage compensation strategy to generate the target AGV control instructions.

[0024] The execution unit is used to repeatedly execute the content acquired from the acquisition unit to the fusion unit to form closed-loop control until the AGV completes the tracking task of the entire fixed route.

[0025] The present invention also provides a computer device, the computer device including a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the above-described method.

[0026] The present invention also provides a storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0027] The advantages of this invention compared to existing technologies are as follows: By constructing an offline floor environment model containing data on slippery areas in a real factory, and combining it with a fixed-route tracking environment model to form a high-fidelity training scenario, the deep reinforcement learning algorithm can fully learn the vehicle dynamic response characteristics under different road surface types (such as slippery steel plates, oily surfaces, ramp joints, etc.). The resulting fixed-route tracking strategy library and slippage compensation strategy library have strong generalization capabilities, reducing path deviation and attitude oscillations caused by sudden changes in road surface adhesion. Furthermore, by introducing a road surface type recognition mechanism based on real-time road surface image / texture features, combined with the AGV's current motion state and load information, an initial compensation strategy is quickly matched from the slippage compensation strategy library, and further fine-tuned online according to the actual motion response to form a dynamic slippage compensation strategy highly adapted to the current working conditions. This mechanism transforms from "passive response" to "active prediction + adaptive adjustment," effectively suppressing drive wheel slippage, sideslip, and vehicle instability. Furthermore, by fusing basic path tracking control commands with dynamic slippage compensation control commands across multiple objectives, the system not only ensures high-precision tracking performance along the preset route but also enhances safety redundancy under sudden low-adhesion conditions without sacrificing task efficiency. This avoids the strategic conflict between "track preservation" and "slippage prevention" in traditional methods. Moreover, the hybrid intelligent control paradigm of "offline training + online fine-tuning" balances algorithm training efficiency with on-site adaptability. It eliminates the need for costly real-time high-precision maps or complex physical modeling, making it suitable for scenarios with frequently occurring heterogeneous road surfaces and dynamic load changes in typical AGV operating environments such as factories, warehouses, and ports. This provides a feasible technical path for the stable operation of large-scale AGV clusters under high safety and high reliability requirements.

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0029] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1 This is a schematic diagram illustrating an application scenario of the anti-skid control method for fixed routes on complex road surfaces provided in an embodiment of the present invention.

[0031] Figure 2 A flowchart illustrating an anti-skid control method for a fixed route on a complex road surface provided in an embodiment of the present invention;

[0032] Figure 3A schematic block diagram of an anti-skid control device for a fixed route on a complex road surface provided in an embodiment of the present invention;

[0033] Figure 4 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0035] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0036] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0037] Please see Figure 1 and Figure 2 , Figure 1 This is a schematic diagram illustrating an application scenario of the anti-skid control method for fixed routes on complex road surfaces provided in this embodiment of the invention. Figure 2 This is a schematic flowchart illustrating an anti-slip control method for AGVs on fixed routes in complex road conditions, provided in an embodiment of the present invention. This anti-slip control method for AGVs on fixed routes in complex road conditions is applied to a server, which interacts with the terminal. It overcomes the technical bottleneck of performance degradation in complex road environments caused by traditional control methods, improving the control accuracy, operational stability, and environmental adaptability of AGVs when performing fixed-route tasks in real industrial scenarios, and possesses promising prospects for industrial application.

[0038] Figure 2 This is a flowchart illustrating an anti-skid control method for fixed routes on complex road surfaces provided in an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S170.

[0039] S110. A fixed route tracking environment model is constructed based on a preset fixed running route, and a factory floor environment model is constructed based on actual measured data of slippery road surfaces in the actual factory scenario. The two together constitute an offline training environment model.

[0040] Specifically, a fixed route refers to a pre-planned and repeatedly executed closed or semi-closed path. First, based on the pre-defined fixed routes (e.g., consisting of a series of waypoints or navigation trajectories) that AGVs need to execute in actual application scenarios such as factories or warehouses, a high-fidelity fixed route tracking environment model is constructed. This model includes path geometry information (such as curvature, slope, and turning radius), speed constraints, task timing, and other elements to simulate the standard tracking behavior of AGVs under ideal conditions. Simultaneously, based on long-term measured data of easily slippery areas collected from the target factory site (including but not limited to physical parameters such as friction coefficients, texture images, humidity, and temperature of different floor materials such as epoxy flooring, steel plates, tiles, joints, oil stains, and water accumulation areas), a factory floor environment model is constructed. This model labels the road surface type and its corresponding adhesion characteristics in map form for each area and can be spatially registered with the AGV's operating path. These two sub-models together constitute an offline training environment model that integrates path tasks and real ground disturbances, providing a simulation platform that closely resembles actual working conditions for subsequent policy learning.

[0041] In other words, by coupling theoretical paths with real ground disturbance data in modeling, the realism and diversity of the training environment are improved, enabling the control strategies obtained from subsequent training to have prior knowledge of complex road surfaces, thus avoiding the technical problem of overfitting to ideal conditions in a pure simulation environment.

[0042] In one embodiment, the construction of a fixed-route tracking environment model based on a preset fixed running route and the construction of a factory floor environment model based on measured data of slippery surfaces in actual factory scenarios, together constituting an offline training environment model, include:

[0043] Acquire the preset fixed running route data of AGV, including the path point coordinate sequence, curvature change information and slope distribution, and construct a fixed route tracking environment model;

[0044] Specifically, the first step is to obtain data on the pre-defined, fixed operating routes that the AGVs need to repeatedly execute in the target factory or warehouse environment. This data is typically provided by a higher-level scheduling system or path planning module, and includes:

[0045] Path point coordinate sequence: A sequence of two-dimensional or three-dimensional spatial points represented with high precision (such as centimeters), which constitutes the geometric trajectory that the AGV needs to strictly track;

[0046] Curvature variation information: By differentiating the path points or fitting spline curves, the instantaneous curvature or turning radius of each road segment is calculated to characterize the curve characteristics of the path.

[0047] Slope distribution data: Combining the digital elevation model (DEM) of the factory area or the results of lidar scanning, the longitudinal and transverse slope values ​​along the path are marked to reflect the influence of gravity on vehicle dynamics.

[0048] Based on the structured data described above, a fixed-route tracking environment model is constructed in a simulation platform (such as Gazebo, CARLA, or a self-developed physics engine). This model not only reproduces the path geometry but also embeds kinematic / dynamic constraints (such as maximum speed, acceleration, and steering angle limits) to simulate the standard tracking behavior of AGVs under ideal ground conditions.

[0049] In other words, by faithfully reproducing the geometric and dynamic characteristics of the actual task path, an accurate task context is provided for subsequent policy training, ensuring that the learned policy can be directly transferred to the real scene and avoiding control mismatch caused by excessive path abstraction.

[0050] In the actual operating area of ​​the factory, oil-stained roads, water-stained roads, gravel-scattered areas, and elevation difference joint areas were selected as typical slippery road sections. Road surface images or texture samples of each section under different lighting conditions were collected, and the wheel slip rate, heading angle fluctuation and trajectory deviation of the AGV when passing through in the empty, half-loaded and full-loaded states were recorded simultaneously to form actual measurement data.

[0051] Specifically, to accurately reflect the complex ground conditions of the factory site, this embodiment manually identifies and delineates four typical slippery road sections within the AGV's regular operating area as key modeling objects:

[0052] Oily road surfaces: commonly found in machining areas or equipment maintenance passages, where there is residual lubricating oil or cutting fluid on the surface;

[0053] Water-stained pavement: often appears in cleaning areas, entrances and exits or areas where water seeps in during rain, and has high reflectivity and low coefficient of friction;

[0054] Areas with scattered sand and gravel: such as around material loading and unloading points, where fine particles accumulate on the ground, resulting in uneven rolling resistance of tires;

[0055] Areas with height differences: such as the joints of different flooring materials (epoxy + steel plate), expansion joints, or threshold transitions, can easily cause wheels to be suspended in the air or to slip due to impact.

[0056] For each type of road segment mentioned above, a large number of road surface images or texture samples are collected using vehicle-mounted cameras or fixed vision sensors under various lighting conditions (such as strong daylight, weak nightlight, backlight, glare, etc.). Simultaneously, high-precision inertial measurement units (IMUs), wheel speed encoders, and RTK-GNSS positioning systems are deployed. When the AGV passes through each road segment under three typical operating conditions (empty load, 0% rated load), the following key dynamic response data are recorded synchronously:

[0057] Wheel slip ratio: Calculated by comparing the theoretical speed of the drive wheel with the actual ground speed, used to quantify the degree of slippage;

[0058] Heading angle fluctuation: reflects the attitude stability of the vehicle body under lateral disturbances;

[0059] Trajectory deviation: This refers to the lateral deviation between the actual trajectory and the preset path, representing the degree of tracking performance degradation.

[0060] In other words, through systematic field measurements of multiple working conditions, multiple environments, and multiple indicators, the nonlinear coupling relationship between "road surface type - load state - vehicle response" is fully captured, laying a data foundation for building a physically meaningful floor environment model and improving the model's ability to represent real slippage behavior.

[0061] Based on measured data, a factory floor environment model was established, which includes road surface type labels, load levels and the mapping relationship between slippage behavior.

[0062] Specifically, based on the aforementioned measured data, each group of samples is structurally labeled to form a dataset containing the following elements:

[0063] Road surface type labels (such as composite labels like "oil stains-steel plate" or "water stains-epoxy");

[0064] Load level indicator (no load / half load / full load);

[0065] The slip behavior feature vector (including peak slip ratio, standard deviation of heading angle, maximum deviation of trajectory, etc.).

[0066] Furthermore, machine learning methods (such as cluster analysis, regression mapping, or graph neural networks) can be used to establish a mapping model between road surface type and skidding behavior. For example, a conditional probability model P(skid response | road surface type, load) can be constructed, or a feedforward neural network can be trained with road surface image features and load parameters as inputs, and expected slip rate and stability risk score as outputs. This model constitutes the factory floor environment model, which is essentially a "skid knowledge base" that integrates perception and dynamic priors.

[0067] In other words, by transforming discrete measured phenomena into a computable and queryable structured environmental model, the control system can predict potential slippage risks based on the current perceived input, thus achieving a leap from "experience-driven" to "model-driven".

[0068] The fixed route tracking environment model is integrated with the factory floor environment model to form an offline training environment model.

[0069] Specifically, the aforementioned fixed-route tracking environment model and the factory floor environment model are spatially aligned and logically coupled:

[0070] In the simulation environment, each coordinate point along the preset path is associated with its corresponding road surface type label (matched according to the measured map); when the AGV simulation body runs to a certain road segment, the slip behavior model corresponding to that road segment is automatically activated, and the corresponding friction coefficient disturbance or random slip noise is injected into the underlying dynamics engine; at the same time, according to the current simulation load state, the matching slip response characteristics are called to make the simulation closer to the real physical performance.

[0071] The resulting offline training environment model retains the accuracy of the task path while incorporating the complex disturbance characteristics of a real factory floor, making it a high-fidelity training ground that combines task orientation with environmental realism.

[0072] In other words, this ensemble model effectively bridges the "reality gap" between simulation and reality, enabling control policies trained within it to possess strong generalization capabilities and a high deployment success rate. Compared to traditional training environments that only use ideal road surfaces or random noise disturbances, this model can improve the adaptability of policies to specific industrial scenarios, providing high-quality, highly relevant training data sources for subsequent deep reinforcement learning.

[0073] S120, in the offline training environment model, uses deep reinforcement learning algorithm to train and generate a fixed route tracking strategy library and a slippage compensation strategy library respectively.

[0074] Specifically, in the completed offline training environment model, Deep Reinforcement Learning (DRL) algorithms (such as PPO, SAC, or DDPG) are used to train two types of policies respectively:

[0075] Fixed route tracking strategy library: Using path tracking error (such as lateral deviation, heading angle deviation), speed tracking accuracy, energy consumption, etc. as reward functions, it learns the optimal control strategy to achieve high-precision trajectory tracking on undisturbed or standard road surfaces;

[0076] Slippage compensation strategy library: With the goal of suppressing slippage (such as abnormal wheel speed difference, sudden change in yaw rate, and increase in lateral displacement), it combines different road surface type labels as part of the state input to learn active compensation actions (such as torque distribution adjustment, speed limit, steering fine-tuning, etc.) for specific low-adhesion scenarios.

[0077] During training, the DRL agent performs tens of thousands to millions of interactions in randomized scenarios that include various road surface changes, load variations, and initial disturbances. Eventually, it converges to form a structured policy library, where each policy can be indexed to a specific road surface type-load-speed combination.

[0078] In other words, by decoupling and optimizing the "tracking" and "anti-slip" tasks respectively, the baseline performance of path tracking is preserved, while the system is given a specific ability to deal with slippage risks. The structured design of the strategy library supports fast retrieval and online invocation, laying the foundation for real-time control.

[0079] In one embodiment, the step of training a fixed-route tracking strategy library and a slippage compensation strategy library using a deep reinforcement learning algorithm in the offline training environment model includes:

[0080] In the offline training environment model, with the goal of minimizing the lateral deviation and heading error between the actual AGV trajectory and the preset route, a fixed route tracking strategy library is generated by training a deep reinforcement learning algorithm.

[0081] Specifically, in the already constructed offline training environment model, deep reinforcement learning training for path tracking capability is first conducted, focusing on the core requirement of stable AGV operation along a preset path. The specific implementation process is as follows:

[0082] The virtual agent used for training takes the real-time status of the AGV in the simulation environment as input. This status information includes the vehicle's current position coordinates, driving direction, forward speed, lateral distance deviation from the target path, angular deviation between the vehicle's heading and the path tangent, and the curvature of the path ahead. By integrating this information, the agent can comprehensively perceive its deviation from the ideal trajectory. The control actions output by the agent are instructions that can be directly applied to the AGV's underlying actuators, such as the steering angle of the front wheels, the target torque or acceleration command of the drive motor, and the specific form is adapted according to the actual drive structure of the AGV (such as differential drive, steering wheel drive, etc.).

[0083] During training, the system's optimization objective is to minimize the lateral positional and heading angle deviations between the AGV's actual trajectory and the preset route. To this end, a composite reward mechanism was designed: a higher positive reward is given when the AGV gets closer to the target path and its heading is more aligned; a negative penalty is applied when there is a significant lateral deviation or heading misalignment. Furthermore, to improve driving smoothness, a term to suppress drastic changes in steering action is introduced, and speed tracking accuracy and energy consumption are moderately guided. If the AGV deviates from the path beyond the safety tolerance during simulation (e.g., lateral deviation greater than 10cm), the current training iteration is immediately terminated and a penalty is applied to strengthen the strategy's ability to identify and adapt to safety boundaries.

[0084] Training employs mature deep reinforcement learning algorithms (such as the soft Actor-Critics algorithm or the proximal policy optimization algorithm) and involves extensive interactive learning in an offline training environment. This environment not only replicates the geometry of the real path but also incorporates subtle perturbations (such as small gradient changes or air resistance) to enhance the policy's robustness. After thorough training, a neural network policy is ultimately obtained that can autonomously output high-precision tracking control commands based on the current state.

[0085] The strategy is solidified and stored, and then classified and cataloged according to path characteristics (such as straight road segments, large curvature curves, continuous S-curves, etc.) or operating speed ranges to form a fixed route tracking strategy library with a clear structure and easy access.

[0086] In other words, the strategy library obtained through the above training method can adaptively handle nonlinear dynamic problems in path tracking without relying on an accurate vehicle dynamics model. Compared with traditional proportional-integral-derivative controllers or model predictive controllers, this strategy exhibits stronger stability and response speed on complex road sections such as curves and slope changes. It can stably control the lateral tracking error of the AGV within 1.5 cm without significant overshoot or oscillation, thus improving the accuracy and reliability of task execution.

[0087] In the offline training environment model, with the goals of suppressing wheel slippage, reducing attitude vibration, and maintaining driving stability, corresponding slippage compensation strategies are trained and generated for each road surface type and load combination to form a slippage compensation strategy library.

[0088] Specifically, within the same offline training environment model, specialized training for anti-skid and driving stability is conducted simultaneously to generate a skid compensation strategy library. The implementation details are as follows:

[0089] First, based on the typical road surface types (such as oily surfaces, water-stained areas, gravel-strewn areas, and joints with elevation differences) already labeled in the factory floor environment model, as well as the common load states of AGVs (empty, half-loaded, and fully loaded), the entire training task is divided into multiple independent sub-scenes. Each sub-scene corresponds to a specific "road surface type and load combination", such as "oily road surface with a fully loaded state" or "water-stained area with an empty state".

[0090] In each sub-scenario, the state input used for training includes not only basic vehicle motion information (such as speed, acceleration, yaw rate, etc.), but also explicitly includes the current road surface type and load level information. This design allows the trained strategy to dynamically adjust its compensation behavior according to specific operating conditions.

[0091] The strategy outputs not complete control commands, but a set of compensation quantities to correct the main control commands. These compensation quantities include torque adjustment offset between the left and right drive wheels, maximum allowable acceleration limit, steering system damping gain adjustment value, and overall speed attenuation ratio. These compensation quantities will be integrated with the basic commands output by the main path tracking strategy during operation to achieve fine-tuning of slippage risk, rather than completely replacing the original control logic.

[0092] The training optimization focuses on three aspects: first, effectively suppressing wheel slippage, for example, by reducing the speed difference between the drive wheels to decrease slip; second, reducing vehicle posture vibration during driving, especially high-frequency yaw; and third, maintaining overall driving stability to avoid loss of control due to sudden low adhesion. To this end, the reward mechanism emphasizes encouraging behaviors with low slippage, gentle yaw changes, and controlled lateral movement, while imposing severe penalties for severe slippage or posture instability. In addition, a moderate path tracking requirement is retained to prevent compensatory actions from excessively interfering with the main task.

[0093] For each road surface and load combination, the system independently completes a full round of deep reinforcement learning training, ultimately generating a lightweight, dedicated compensation strategy. All strategies are organized and stored according to their corresponding working condition labels, forming a slippage compensation strategy library that supports rapid retrieval of matching items based on real-time identification results during actual operation.

[0094] In other words, this scenario-based, finely tuned training approach enables each compensation strategy to accurately address the slippage mechanism under specific operating conditions. For example, when passing through an oil-slicked area under full load, the strategy will preemptively limit acceleration intensity and evenly distribute driving force between the left and right wheels; while when crossing a joint area under no load, it will focus on suppressing high-frequency vibrations and directional jitter caused by road impacts. Compared to traditional methods that use a uniform anti-slip logic, this solution can more effectively suppress slippage without compromising path tracking accuracy due to excessive intervention.

[0095] More specifically, offline training is for independently training the "fixed route tracking strategy" and the "slippage compensation strategy," generating two basic policy libraries to provide a benchmark for subsequent online control. This offline training mainly consists of four parts (environment modeling, fixed route tracking strategy training, slippage compensation strategy training, and optimal policy output):

[0096] Offline training environment model construction: ① Fixed route tracking environment model: Based on the AGV's preset fixed route, it includes the route's coordinate information, curvature changes, slope data, and tracking error data of the AGV under different driving speeds and load states on a flat road surface. The core is used to train the fixed route tracking capability; ② Factory floor environment model: The core of the model includes the measured and calibrated slip coefficients of typical slippery road surfaces in the factory scene, including oily road surfaces, water-stained road surfaces, gravel road surfaces, and road surfaces with elevation differences and joints. At the same time, it integrates the full load state data of AGVs such as empty, half-loaded, and full-loaded, as well as the slip behavior characteristics of AGVs under the corresponding road surfaces and historical control parameter adjustment data;

[0097] Fixed-route tracking strategy training: A tracking strategy model is constructed based on the PPO (Proximal Policy Optimization) reinforcement learning algorithm, with the core objective of "minimizing the fixed-route tracking error". A dedicated reward function R is designed: R = 1 - exp(-k·e), where e is the lateral / longitudinal deviation between the actual AGV trajectory and the fixed route, and k is a coefficient; the smaller the tracking error e, the higher the reward value R, driving the model to learn the optimal speed control strategy. The model is trained offline iteratively using a tracking strategy dataset until the model converges (the tracking error stabilizes at the centimeter level), generating a fixed-route tracking strategy library containing the optimal tracking control parameters (linear velocity and angular velocity) under different driving speeds and loads.

[0098] Slippage Compensation Strategy Training: A slippage compensation strategy model is constructed based on the PPO reinforcement learning algorithm. The core objective is to "minimize the AGV slippage deviation and heading difference fluctuation" to achieve adaptive anti-slip control under different road surface and load conditions. ① Definition of State Space and Action Space: State Space: The input dimension includes the core parameters of the factory floor environment model, specifically the actual measured slippage coefficient of the road surface, the real-time load status of the AGV (empty / half-loaded / full-loaded), the wheel speed difference of the drive wheels, position, heading angle deviation, and current travel speed; Action Space: The output dimension is the AGV anti-slip compensation control parameters (adaptive adjustment coefficient of travel speed). ② Dedicated Reward Function Design: A multi-objective weighted reward function Rtotal=w1·R1+w2·R2+w3·R3 is designed to balance anti-slip effect, path tracking accuracy, and operational efficiency. Here: R1=exp(-α·λ); Slippage suppression reward term, where λ is the real-time slippage coefficient and α is the weighting coefficient; the smaller the slippage coefficient, the higher the R1 reward value, driving the model to learn control strategies to reduce slippage; R2=1-exp(-β·Δv); Heading stability reward term, where Δv is the deviation between the AGV's real-time wheel speed and the theoretical wheel speed, and β is the weighting coefficient; the smaller the wheel speed deviation, the higher the R2 reward value; R3=γ·vsafe; Safe speed reward term, where vsafe is the safe driving speed under the current operating conditions, and γ is the weighting coefficient; under the premise of ensuring anti-slip and heading stability, it incentivizes the model to maintain a reasonable driving speed, avoiding excessive speed reduction that affects efficiency; w1, w2, and w3 are normalized weighting coefficients that satisfy w1+w2+w3=1, and can be dynamically adjusted according to the actual operating conditions of the factory. ③ Offline Iterative Training and Strategy Library Generation: Using a full-condition dataset of the factory floor environment model (covering 4 typical slippery road surfaces and 3 load states), the strategy model is trained offline iteratively. In each training round, the model outputs compensating control actions based on the input state parameters, optimizing the strategy network parameters through environmental feedback rewards. Training stops when the model converges (slippery coefficient fluctuation stabilizes within ±5%, wheel speed deviation is controlled within ±2%). Finally, a slippery compensation strategy library is generated, which stores the optimal compensation control parameter combinations categorized by "road surface type - load state," providing precise anti-slip strategy support for subsequent online control.

[0099] Strategy Output: The trained and mature fixed route tracking strategy library and the multi-road slip coefficient adaptation strategy library (i.e. slip compensation strategy library) are synchronously stored in the system. The fixed route tracking strategy library is directly imported into the strategy fusion module as the basic strategy, and the multi-road slip coefficient adaptation strategy library is imported into the online slip compensation adjustment module as the adjustment benchmark.

[0100] S130: Obtain task information of AGV executing a fixed route, and collect in real time the road surface image or texture features of the current AGV location, AGV's own motion status information, and load information;

[0101] Specifically, when the AGV performs a fixed-route task, the system simultaneously acquires three types of key information:

[0102] Task information: the currently executing route ID, the sequence of target path points, the planned speed curve, etc.;

[0103] Environmental perception information: Real-time acquisition of road surface images or texture features (such as grayscale gradient, spectral features, and semantic features extracted by CNN) in front of and below the AGV through vehicle-mounted cameras, laser texture sensors, or visual SLAM modules.

[0104] Body status information: including the rotational speed of each drive wheel, acceleration and angular velocity measured by the IMU, motor current / torque, and center of gravity height estimation (calculated from the load weight and distribution), etc.

[0105] In other words, multimodal perception fusion provides sufficient input for accurately identifying the current road surface condition and vehicle dynamics, overcoming the limitations of a single sensor under complex lighting or occlusion conditions.

[0106] In one embodiment, the step of acquiring task information of the AGV executing a fixed route and collecting in real time road surface images or texture features at the current location of the AGV, the AGV's own motion state information, and load information includes:

[0107] Receive the fixed route task instructions currently being executed by the AGV, determine the target path point sequence and the start and end range of the task, and obtain task information;

[0108] Specifically, when an AGV begins performing a transportation or handling task, the host scheduling system (such as a Warehouse Management System (WMS) or Manufacturing Execution System (MES) first issues a fixed-route task instruction to the AGV controller. This instruction explicitly specifies the preset path identifier corresponding to this task, such as "from charging station in area A to workstation 3 on assembly line in area B". Based on this identifier, the controller retrieves the complete sequence of target path points from its locally stored path database, including the precise coordinates of the path start point, intermediate key navigation points, and end point, while also obtaining the start and end range of the task (e.g., only the first half of the path is used for temporary stops). This information together constitutes the task information required for the current operation of the AGV, providing a task context for subsequent path tracking and strategy invocation.

[0109] In other words, by receiving and parsing task instructions in a structured manner, AGVs can accurately identify their current task stage and target trajectory, avoiding misoperations caused by path confusion or incorrect task switching. At the same time, clear task boundaries help the system preload relevant control strategies (such as compensation plans for specific curves or slippery sections), improving response efficiency.

[0110] Real-time acquisition of road surface images or texture features in front of and below the AGV using onboard cameras or vision sensors;

[0111] Specifically, to perceive the ground conditions of the current driving area, the AGV is equipped with an onboard camera or dedicated vision sensor, typically installed at the lower front of the vehicle or in the middle of the chassis, covering a field of view of 0.5 to 2 meters in front of the vehicle and the area directly below it. During operation, the sensor continuously collects road surface images or texture features at a frequency of more than ten frames per second. The collected data can be raw RGB images, grayscale images, or texture descriptors (such as local binary patterns, gradient orientation histograms, etc.) extracted by a front-end image signal processor. This data can effectively reflect visual cues such as ground material, the presence of liquid reflections, scattered particles, and height differences between seams.

[0112] In industrial environments with complex lighting conditions (such as strong backlighting, dimly lit workshops, and stroboscopic light sources), the system can employ multispectral imaging, infrared-assisted illumination, or adaptive exposure control technology to ensure stable and reliable image quality. Furthermore, some embodiments may integrate data from a laser texture scanner to enhance the ability to identify minute bumps or oil films under non-visible light.

[0113] In other words, by enabling real-time, high-frequency road surface visual perception, the AGV gains the ability to "see the road," allowing it to identify risks before entering potentially low-adhesion areas. Compared to traditional methods that rely on post-slip detection, this solution achieves feedforward risk prediction, providing a valuable time window for the early activation of subsequent slip compensation strategies and enhancing the system's proactive safety.

[0114] The AGV's position, speed, heading angle, left and right wheel speed difference, and acceleration information are acquired in real time through an inertial measurement unit, wheel speed encoder, and positioning module to form the AGV's own motion state information.

[0115] Specifically, AGVs integrate multiple onboard sensors to monitor their dynamic behavior in real time, forming comprehensive information about their own motion status. This includes:

[0116] Inertial Measurement Unit (IMU): Installed near the vehicle's center of gravity, it is used to measure the vehicle's three-dimensional acceleration and angular velocity at high frequency, and then calculate key attitude parameters such as the rate of change of heading angle and lateral acceleration.

[0117] Wheel speed encoders: Installed on the left and right drive wheel motor shafts respectively, they accurately record the real-time speed of each wheel. By comparing the speed difference between the left and right wheels, it can be determined whether there is slippage on one side or abnormal steering.

[0118] Positioning module: It can use laser SLAM, QR code navigation, magnetic nail positioning or UWB fusion positioning to provide the AGV's centimeter-level position coordinates and absolute heading angle in the global map.

[0119] The aforementioned sensor data is time-synchronized and filtered and fused (such as by Kalman filtering) by the vehicle controller to generate a stable, low-noise motion state information stream, including current position, forward speed, vehicle heading angle, difference in speed between left and right wheels, longitudinal and lateral acceleration, etc.

[0120] In other words, the fusion of multi-source motion state information not only improves the accuracy and robustness of state estimation but also effectively identifies early signs of slippage. For example, when the speed difference between the left and right wheels suddenly increases while the heading angle does not change as expected, it can be determined that the drive wheels are slipping; when the lateral acceleration increases abnormally without steering input, it may indicate a risk of sideslip. These fine-grained dynamic characteristics provide key basis for the precise triggering and online fine-tuning of slippage compensation strategies.

[0121] The real-time load information of the AGV is obtained through weighing sensors or motor current estimation modules.

[0122] Specifically, the load status of an AGV directly affects its center of gravity position, tire contact pressure, and slippage sensitivity. Therefore, the system obtains real-time load information through the following methods:

[0123] Integrate load cells (such as strain gauge or piezoelectric sensors) into the bottom of the forks or platform to directly measure the weight of the cargo and estimate the center of gravity offset in combination with the cargo position.

[0124] Alternatively, in a low-cost configuration without weighing hardware, a motor current estimation module can be used to deduce the current total load mass by monitoring the steady-state current value of the drive motor during constant speed driving or climbing, combined with known vehicle unloaded parameters. This method is based on the physical principle that "the greater the load, the higher the driving torque required to maintain the same speed, and the greater the motor current."

[0125] The resulting load information is typically output in discrete levels (such as no load, half load, full load) or continuous values ​​(such as kilograms), and together with task information and road surface recognition results, it serves as a key input for strategy matching.

[0126] In other words, by introducing load information, the system can distinguish between different instability mechanisms such as "light load, easy to bounce" and "heavy load, easy to slip," thereby invoking more suitable compensation strategies. For example, under full load, more attention is paid to the balanced torque distribution to prevent drive wheel slippage, while under no-load conditions, the focus is on suppressing attitude vibrations caused by high-frequency vibrations. This load sensing capability improves the targeting and effectiveness of anti-slip control.

[0127] S140 identifies the current road surface type based on road surface images or texture features, and combines the AGV's own motion state information and load information to match the corresponding initial slippage compensation strategy from the slippage compensation strategy library;

[0128] Specifically, based on the collected road surface images / texture features, a pre-trained lightweight classification model (such as MobileNet or a small Transformer) identifies the current road surface type (e.g., "dry epoxy flooring," "wet steel plate," "oil-stained joints," etc.) in real time. Then, combining the current AGV's motion state (e.g., speed, acceleration, yaw rate) and load information (total mass, center of gravity shift), the nearest neighboring strategy entry is retrieved from the slippage compensation strategy library as the initial slippage compensation strategy. This strategy outputs a set of preliminary compensation control quantities, such as the left and right wheel torque difference correction value, the maximum allowable acceleration limit, and the steering gain adjustment coefficient.

[0129] In other words, by achieving a rapid mapping from "perception" to "decision-making," online learning from scratch is avoided, response latency is significantly reduced, and the millisecond-level control cycle requirement of AGVs is met.

[0130] In one embodiment, the step of identifying the current road surface type based on road surface images or texture features, and combining this with the AGV's own motion state information and load information to match a corresponding initial slippage compensation strategy from the slippage compensation strategy library includes:

[0131] The collected road surface images or texture features are classified and identified to determine whether the current road surface belongs to oil stains, water stains, gravel, joints, or normal road surface, so as to obtain the current road surface type;

[0132] Specifically, during AGV operation, the system continuously receives road surface images or texture features from the areas in front and below, collected by onboard cameras or vision sensors. This raw data is fed into a pre-trained lightweight road surface classification model for real-time analysis. This model is typically based on a convolutional neural network architecture (such as MobileNet, EfficientNet, etc.) and has been trained offline using a large number of labeled samples. It can accurately distinguish several typical ground conditions commonly found in industrial environments, including: oily surfaces (reflective surface, dark color, with flow marks), water-stained surfaces (highly reflective, blurred edges), areas with scattered sand and gravel (strong granularity, rough texture), areas with elevation differences and joints (with obvious lines or step-like structures), and normal dry surfaces (uniform texture, no abnormal reflections or debris).

[0133] The classification process is executed multiple times per second, outputting the most likely road type label corresponding to the current vehicle's location. To improve robustness, the system can also perform sliding window smoothing on the recognition results of multiple consecutive frames to avoid misjudgments caused by momentary occlusion (such as forklift shadows or people passing by). In addition, when there are drastic changes in lighting or a decrease in image quality, historical location information and a pre-stored factory ground map can be used for auxiliary verification to ensure the reliability of the recognition results.

[0134] In other words, through high-precision, low-latency road surface type identification, the system can promptly identify the nature of the risk before the AGV enters a potentially low-adhesion area, providing accurate environmental context for subsequent strategy invocation. Compared to traditional methods that rely on post-event skid detection, this feedforward perception mechanism enhances the system's proactive intervention capability and is a key prerequisite for achieving "prevention-oriented" anti-skid control.

[0135] By combining the current road surface type, real-time load information and the current motion status information of the AGV, the slippage compensation strategy library is queried to obtain the initial slippage compensation strategy with the highest matching degree.

[0136] Specifically, after obtaining the current road surface type, the system further integrates real-time load information (such as no load, half load or full load) and AGV's own motion status information (including current speed, acceleration, difference in left and right wheel speeds, rate of change of heading angle, etc.) to form a multi-dimensional query condition.

[0137] This query condition is used to access the slippage compensation strategy library—a structured collection of strategies that were built during the offline training phase. Each strategy in the library is associated with a specific combination of "road surface type + load level + motion state range". For example, "oil-stained road surface + full load + medium speed driving" corresponds to a set of specially optimized compensation parameter configurations.

[0138] The system uses an indexing mechanism to quickly traverse the strategy library and find a strategy entry that perfectly matches the current operating condition. If such an entry exists, it is directly used as the initial slippage compensation strategy. This strategy includes a set of preset compensation control variables, such as the torque adjustment values ​​of the left and right drive wheels, the maximum allowable acceleration limit, the steering damping gain coefficient, and the speed attenuation ratio, which are used to finely correct the underlying execution actions without changing the main path tracking command.

[0139] In other words, by accurately matching multi-dimensional operating conditions, the compensation strategy is ensured to be highly adapted to the current physical scenario, avoiding performance losses caused by a one-size-fits-all approach. For example, when passing through an oil-stained area under full load, the strategy will focus more on limiting the driving force to prevent slippage; while when passing through a joint area under no load, it will prioritize suppressing high-frequency vibrations. This scenario-customized compensation improves the effectiveness and economy of control.

[0140] If no perfect match is found, the strategy with the closest load under the same road surface type is selected as the initial slip compensation strategy.

[0141] Specifically, in actual operation, there may be situations where the strategy library does not contain a strategy entry that is completely consistent with the current operating conditions. For example, the current load is 70% of the rated load, but the strategy library only contains three preset levels: no load (0%), half load (50%), and full load (100%).

[0142] To address such situations, the system employs an approximate matching mechanism: First, it identifies the road surface type that matches the current identification result (e.g., "waterlogged road surface"). Then, among all strategies within that type, it selects the one with the load level closest to the current actual load as the alternative. For example, if the current load is 70%, the "full load" strategy is preferred over "half load" because its physical characteristics (e.g., ground pressure, inertia) are closer to the actual situation.

[0143] To further improve adaptability, the system can also perform linear interpolation or scaling fine-tuning on the output of the selected strategy (such as adjusting the torque compensation amplitude according to the load ratio), but this step belongs to the subsequent online fine-tuning stage, and this stage only completes the rapid selection of the initial strategy.

[0144] In other words, this fault-tolerance mechanism effectively resolves the contradiction between the limited coverage of the strategy library and the continuous changes in actual operating conditions, ensuring that the system has an available initial compensation scheme under any load state and avoiding control failure due to missing strategies. At the same time, by prioritizing the consistency of road surface type and then optimizing the load matching degree through a layered matching logic, the physical rationality of the strategy is preserved to the greatest extent, ensuring basic anti-skid performance.

[0145] S150, based on the real-time acquisition of the AGV's actual motion response and road surface image or texture features, fine-tunes the initial slippage compensation strategy online to generate a dynamic slippage compensation strategy that adapts to the current working conditions.

[0146] Specifically, based on the initial strategy, the system continuously monitors the actual motion response of the AGV (such as whether there are sudden changes in wheel speed, increased sideslip angle, accelerated deviation from the trajectory, or other signs of slippage), and compares it with the expected response. If the detected deviation exceeds the threshold, an online fine-tuning mechanism is activated.

[0147] By utilizing lightweight online learning modules (such as the idea of ​​Model Reference Adaptive Control (MRAC) or incremental reinforcement learning to update strategy parameters), and taking the current road surface features and motion state as context, the output of the initial compensation strategy is slightly adjusted to generate a dynamic slippage compensation strategy that better fits the instantaneous working conditions.

[0148] In other words, by compensating for the shortcomings of offline policy libraries in covering all on-site details, personalized adaptation of "a thousand road conditions, a thousand policies" can be achieved, thereby improving the robustness of the system under unknown or unmodeled road surfaces (such as sudden watering or temporary paving).

[0149] In one embodiment, the online fine-tuning of the initial slippage compensation strategy based on real-time acquired AGV actual motion response and road surface image or texture features to generate a dynamic slippage compensation strategy adapted to the current working conditions includes:

[0150] The initial slippage compensation strategy is used as the baseline strategy for online fine-tuning;

[0151] Specifically, after the initial slippage compensation strategy is matched, it is not used directly as the final control command, but is set as the starting point or benchmark strategy for online fine-tuning. This benchmark strategy includes a set of preset compensation control parameters, such as the torque distribution ratio between the left and right drive wheels, the maximum allowable acceleration limit, the steering system damping coefficient, and the speed decay factor. These parameters are optimized for specific "road surface type-load" combinations during the offline training phase and represent the prior optimal solution for that type of working condition.

[0152] In other words, by using offline strategies as the starting point for online adjustments, the high latency and high risk of learning from scratch are avoided, while retaining the basic adaptability to the current scenario, providing a reliable foundation for subsequent fine-tuning.

[0153] Continuously collect the actual motion response of the AGV, including the difference in rotational speed between the left and right drive wheels, the yaw rate of the vehicle body, and the trend of trajectory deviation;

[0154] Specifically, during AGV task execution, the system continuously collects AGV's actual motion response data at a high frequency (typically every 10 to 50 milliseconds) to determine the effectiveness of the current compensation strategy. Key monitoring indicators include:

[0155] Difference in rotational speed between the left and right drive wheels: If one side of the wheel is significantly faster than the other side without any steering input, it may indicate that the wheel on that side is slipping.

[0156] Yaw rate: Reflects how fast the vehicle rotates around its vertical axis; an abnormally high yaw rate usually indicates sideslip or loss of steering control.

[0157] Trajectory deviation trend: By comparing the real-time positioning position with the preset path, the rate of change of lateral deviation is calculated. If the deviation increases rapidly, it indicates that the tracking stability is deteriorating.

[0158] These dynamic response data accurately reflect the actual behavior of AGVs under current ground conditions and control commands, and serve as a direct basis for evaluating the effectiveness of the strategy.

[0159] In other words, by monitoring motion states in multiple dimensions and at high frequency, the system can capture early signs and evolutionary trends of slippage, rather than relying solely on passive responses after severe loss of control, thereby enabling more precise and timely intervention.

[0160] The actual motion response of the AGV, road surface image or texture features, and the baseline strategy are all input into a lightweight online reinforcement learning module.

[0161] Specifically, to achieve real-time adaptive adjustment, the system integrates a lightweight online reinforcement learning module. This module has a simple structure and low computational overhead, making it suitable for deployment on the limited computing power platform of the AGV onboard controller. Within each control cycle, this module receives three types of key inputs:

[0162] The actual motion response of the AGV (such as the speed difference, yaw rate, and trajectory deviation trend mentioned above);

[0163] Current road surface image or texture features (used to confirm whether the ground condition has changed, such as transitioning from water stains to dry areas);

[0164] The current baseline strategy being used (i.e., the parameter configuration of the initial slippage compensation strategy).

[0165] These three types of information together constitute a complete description of the closed loop of "current control behavior - environmental state - system response".

[0166] In other words, by fusing information from multiple sources, the online learning module not only focuses on "whether the result is good or bad," but also understands "why it is bad," thereby making adjustments with more physical meaning. For example, if the road surface has dried but the strategy is still excessively restricting the driving force, the system can determine that the compensation intensity needs to be reduced.

[0167] The online reinforcement learning module evaluates the effectiveness of the current baseline strategy in suppressing slippage and maintaining stability based on the deviation between the current motion response and the desired driving state, in order to obtain the evaluation results;

[0168] Specifically, the core function of the online reinforcement learning module is to evaluate the actual effectiveness of the current baseline strategy in real time. The evaluation logic is based on a preset performance objective: under the current road conditions, whether it effectively suppresses wheel slippage, maintains vehicle stability, and ensures the continuity of trajectory tracking. Specifically, the module will determine:

[0169] Is the difference in speed between the left and right wheels within a reasonable range?

[0170] Whether the yaw rate is stable and without violent oscillations;

[0171] Does the lateral deviation of the trajectory tend to converge rather than diverge?

[0172] If all the above indicators perform well, the current strategy is deemed effective and no adjustment is needed; if one or more indicators deteriorate, the strategy is deemed inadequate and a corresponding effectiveness assessment result is generated, such as qualitative or semi-quantitative feedback such as "excessive driving force compensation", "insufficient steering damping" or "excessive speed decay".

[0173] In other words, this evaluation mechanism based on behavioral feedback enables the system to have the ability to "self-reflect," distinguishing whether strategy failure stems from sudden environmental changes or parameter mismatch, thus providing guidance for subsequent precise adjustments.

[0174] Based on the evaluation results, the control parameters in the baseline strategy are adjusted incrementally to generate a dynamic slippage compensation strategy that adapts to the current instantaneous operating conditions.

[0175] Specifically, based on the above evaluation results, the online reinforcement learning module performs small, gradual adjustments to the key control parameters in the baseline policy, i.e., "incremental adjustments." For example:

[0176] If slippage is detected due to excessive driving force, the torque distribution ratio will be slightly reduced or the acceleration limit will be further tightened.

[0177] If the vehicle body is found to be vibrating at high frequencies due to insufficient damping, the damping gain of the steering system should be increased appropriately.

[0178] If the road surface has regained high adhesion but the strategy is still overly conservative, then gradually relax the speed attenuation factor to restore operating efficiency.

[0179] All adjustments are made in small increments to ensure smooth control actions without inducing new oscillations or instabilities. The adjusted parameter combination constitutes a dynamic slippage compensation strategy adapted to the current instantaneous operating conditions and will be immediately used in the command fusion of the next control cycle.

[0180] In other words, the incremental adjustment mechanism achieves adaptive compensation for unknown disturbances (such as unmodeled road surfaces, sudden watering, and minor load changes) while ensuring system stability. Compared with fixed strategies or drastic strategy switching, this method improves the continuity and robustness of control and avoids secondary instability caused by "overcorrection".

[0181] More specifically, the core objective of online fine-tuning is to achieve real-time dynamic adjustment of the slippage compensation strategy, accurately adapting to the dynamic changes in the actual road surface slippage coefficient, and ensuring the anti-slip stability and control precision during AGV operation. Its specific working logic consists of the following four stages, each linked in a closed loop and progressively advancing:

[0182] Operating condition matching: The system receives key operating condition data such as the current road surface slip coefficient and load status from the sensing module in real time. Through an efficient matching algorithm, it quickly retrieves the multi-road surface slip coefficient adaptation strategy library (i.e. slip compensation strategy library), locates and calls the basic slip compensation strategy that best matches the current operating condition, and provides a benchmark for subsequent fine-tuning.

[0183] Deviation analysis: Using the baseline slip coefficient determined during the offline training phase as a reference, the difference between the slip coefficient detected on the actual road surface and the actual slip coefficient is compared and a risk assessment is performed. If the actual slip coefficient is higher than the baseline value, it indicates that the current road surface anti-skid difficulty has increased and slippage is more likely to occur, requiring targeted strengthening of anti-skid compensation; if the actual value is lower than the baseline value, the compensation parameters can be appropriately optimized to achieve a balance between precise control and efficiency.

[0184] Online fine-tuning: The TD3 (Twin Delayed DDPG) online reinforcement learning algorithm is adopted. With the core optimization goal of "minimizing the current AGV slippage risk and ensuring operational stability", the basic slippage compensation strategy called in the working condition matching process is dynamically fine-tuned. The final output is a dynamic slippage compensation strategy and corresponding parameters that can be directly adapted to the current actual road surface and load conditions.

[0185] Strategy Output: Output the optimal strategy after fine-tuning.

[0186] S160: The basic control instructions output by the fixed route tracking strategy library based on the current path point are merged with the compensation control instructions output by the dynamic slippage compensation strategy to generate the target AGV control instructions.

[0187] Specifically, basic control commands (such as target speed, front wheel angle, and desired acceleration) from the fixed route tracking strategy library are fused with compensation control commands (such as torque bias, speed attenuation factor, and steering damping gain) output by the dynamic slippage compensation strategy, performing multi-objective fusion. The fusion method can employ weighted superposition, priority arbitration, or constraint optimization based on a safety barrier function to ensure that driving stability is prioritized without sacrificing the core objective of path tracking. The final target AGV control command is then output and sent to the underlying drive and steering actuators.

[0188] In other words, by achieving synergistic optimization of the two major goals of "precise tracking" and "active anti-slip", we can avoid path oscillation caused by overcompensation or loss of control caused by ignoring slippage, as is the case with traditional methods, and achieve a balance between safety and task efficiency.

[0189] In one embodiment, the step of fusing the basic control commands output by the fixed route tracking strategy library based on the current path point with the compensation control commands output by the dynamic slippage compensation strategy to generate target AGV control commands includes:

[0190] Based on the current pose of the AGV and the next target path point in the preset fixed route, the basic control instructions for maintaining path tracking are queried from the fixed route tracking strategy library and output. The basic control instructions include basic linear velocity and basic angular velocity.

[0191] Specifically, during AGV operation, the system acquires its current pose information (including position coordinates and vehicle orientation) in real time, and combines this information with a pre-planned sequence of path points in a fixed route to determine the next target path point to be tracked. This target path point represents the ideal position and direction that the AGV needs to approach within the next short time window.

[0192] Based on the geometric relationship between the current pose and the target path point (such as lateral deviation, heading deviation, path curvature, etc.), the system retrieves the best-matching strategy entry from a fixed-route tracking strategy library. This strategy library is trained offline using deep reinforcement learning and is specifically designed to achieve high-precision path tracking under standard or ideal ground conditions. The selected strategy then outputs a set of basic control commands, mainly including basic linear velocity (i.e., the target speed for the vehicle to move forward or backward) and basic angular velocity (i.e., the target turning rate of the vehicle rotating around its vertical axis). These two commands together define the ideal motion behavior that the AGV should perform in a undisturbed environment.

[0193] In other words, by calling a well-trained path tracking strategy, the system can automatically generate smooth and accurate basic motion commands on various road sections such as curves, straight sections, and slope changes, ensuring that the AGV always moves efficiently along the preset trajectory, providing a stable and reliable main control benchmark for subsequent anti-slip compensation.

[0194] The compensation control quantity matching the current operating condition is extracted from the dynamic slippage compensation strategy. The compensation control quantity includes the compensation linear velocity adjustment and the compensation angular velocity correction for suppressing slippage.

[0195] Specifically, the system extracts compensation control quantities that match the current operating state from the dynamic slippage compensation strategy generated in the previous steps (i.e., the strategy adapted to the current instantaneous operating conditions after online fine-tuning). These compensation quantities are not complete control commands, but rather corrections to the basic commands, specifically designed to address slippage risks caused by low-adhesion road surfaces, load changes, or sudden disturbances.

[0196] Specifically, the compensation control amount includes two parts:

[0197] Compensation linear velocity adjustment: This is used to moderately reduce or limit the vehicle's forward speed to minimize the possibility of the drive wheels slipping due to excessive torque. For example, on wet or slippery surfaces, this adjustment may manifest as a downward correction to the base linear velocity.

[0198] Compensation angular velocity correction: Used to fine-tune steering actions and suppress vehicle yaw or oversteer caused by sudden changes in lateral forces. For example, when passing through a seam area, this correction may apply damping to the base angular velocity, slowing down the steering response.

[0199] The magnitude and direction of these compensation amounts are dynamically generated by the dynamic slippage compensation strategy based on real-time perception and evaluation results, and are highly specific to the operating conditions.

[0200] In other words, the design of the compensation control quantity embodies the principle of "precise intervention and minimal disturbance"—applying just the right correction only in the necessary dimensions (speed or steering) to effectively suppress slippage and avoid unnecessary interference with the main task, thus ensuring the precision and efficiency of control.

[0201] According to the preset fusion rules, the base linear velocity and the compensation linear velocity adjustment amount are superimposed to obtain the target linear velocity; at the same time, the base angular velocity and the compensation angular velocity correction amount are superimposed to obtain the target angular velocity.

[0202] Specifically, the system synthesizes the two types of instructions according to preset fusion rules. These fusion rules typically employ a direct superposition method, which is logically simple, computationally efficient, and suitable for vehicle-mounted real-time control systems.

[0203] Add the base linear velocity to the compensation linear velocity adjustment (if the adjustment is negative, it is actually a deceleration) to obtain the final target linear velocity;

[0204] The base angular velocity is added to the compensated angular velocity correction (the correction can be positive or negative, used to enhance or weaken steering actions) to obtain the final target angular velocity.

[0205] This superposition is not a simple arithmetic operation, but a reasonable combination in a physical sense: the basic command represents "how I want to go", the compensation command represents "how the ground conditions require me to adjust", and the two together determine "how I should actually go now".

[0206] In other words, through a clearly structured instruction fusion mechanism, the system achieves an organic unity between the path tracking objective and the anti-skid safety objective. On high-adhesion road sections, the compensation amount is close to zero, and the AGV operates almost entirely according to the basic strategy, ensuring efficiency; on low-adhesion road sections, the compensation amount is automatically activated, prioritizing stability.

[0207] The target linear velocity and target angular velocity are integrated to form the target AGV control command.

[0208] Specifically, the system integrates the obtained target linear velocity and target angular velocity into a complete target AGV control command and sends it to the underlying motion controller (such as motor driver, servo controller, etc.), which then converts it into specific motor speed, torque or steering angle execution signals.

[0209] S170, repeat steps S130 to S160 to form a closed-loop control until the AGV completes the tracking task of the entire fixed route.

[0210] Specifically, the steps S130–S160 described above are executed cyclically at intervals of AGV control cycles (typically 10-100ms), forming a complete perception-decision-execution closed loop. As the AGV moves along a fixed route, the system continuously updates the road surface judgment, strategy matching, and compensation intensity until the task ends.

[0211] In other words, by constructing a dynamic anti-slip control closed loop that covers the entire stroke and all working conditions, the AGV can maintain continuous, stable, and high-precision operation when traversing multiple heterogeneous road surface areas (such as entering a wet and slippery loading and unloading area from a dry workshop).

[0212] The aforementioned anti-slip control method for AGVs on fixed routes in complex road conditions constructs an offline floor environment model containing data on slippery areas in a real factory. This model is then combined with a fixed route tracking environment model to form a high-fidelity training scenario. This allows the deep reinforcement learning algorithm to fully learn the vehicle's dynamic response characteristics under different road surface types (such as slippery steel plates, oily surfaces, and ramp joints). The resulting fixed route tracking strategy library and slippage compensation strategy library possess strong generalization capabilities, reducing path deviation and attitude oscillations caused by abrupt changes in road surface adhesion. Furthermore, by introducing a road surface type recognition mechanism based on real-time road surface images / texture features, combined with the AGV's current motion state and load information, an initial compensation strategy is quickly matched from the slippage compensation strategy library. This initial strategy is then fine-tuned online based on the actual motion response, forming a dynamic slippage compensation strategy highly adapted to the current working conditions. This mechanism transforms from a "passive response" to an "active prediction + adaptive adjustment," effectively suppressing drive wheel slippage, sideslip, and vehicle instability. Furthermore, by fusing basic path tracking control commands with dynamic slippage compensation control commands across multiple objectives, the system not only ensures high-precision tracking performance along the preset route but also enhances safety redundancy under sudden low-adhesion conditions without sacrificing task efficiency. This avoids the strategic conflict between "track preservation" and "slippage prevention" in traditional methods. Moreover, the hybrid intelligent control paradigm of "offline training + online fine-tuning" balances algorithm training efficiency with on-site adaptability. It eliminates the need for costly real-time high-precision maps or complex physical modeling, making it suitable for scenarios with frequently occurring heterogeneous road surfaces and dynamic load changes in typical AGV operating environments such as factories, warehouses, and ports. This provides a feasible technical path for the stable operation of large-scale AGV clusters under high safety and high reliability requirements.

[0213] Figure 3 This is a schematic block diagram of an anti-skid control device 300 for a fixed route on complex road surfaces provided in an embodiment of the present invention. Figure 3As shown, corresponding to the above-described anti-skid control method for fixed routes on complex surfaces of AGVs, the present invention also provides an anti-skid control device 300 for fixed routes on complex surfaces of AGVs. This anti-skid control device 300 includes a unit for executing the above-described anti-skid control method for fixed routes on complex surfaces of AGVs, and the device can be configured in a server. Specifically, please refer to... Figure 3 The anti-skid control device 300 for fixed routes on complex road surfaces used by AGVs includes:

[0214] The construction unit 301 is used to build a fixed route tracking environment model based on a preset fixed running route and a factory floor environment model based on measured data of slippery road surfaces in actual factory scenarios. Together, the two constitute an offline training environment model.

[0215] The generation unit 302 is used to train and generate a fixed route tracking strategy library and a slippage compensation strategy library respectively in the offline training environment model using a deep reinforcement learning algorithm.

[0216] The acquisition unit 303 is used to acquire task information of the AGV executing a fixed route, and to collect in real time the road surface image or texture features of the current AGV location, the AGV's own motion status information, and load information.

[0217] The identification and matching unit 304 is used to identify the current road surface type based on the road surface image or texture features, and to match the corresponding initial slippage compensation strategy from the slippage compensation strategy library by combining the AGV's own motion state information and load information.

[0218] The fine-tuning unit 305 is used to fine-tune the initial slippage compensation strategy online based on the real-time acquired AGV actual motion response and road surface image or texture features, and generate a dynamic slippage compensation strategy that adapts to the current working conditions.

[0219] The fusion unit 306 is used to fuse the basic control instructions output by the fixed route tracking strategy library based on the current path point with the compensation control instructions output by the dynamic slippage compensation strategy to generate the target AGV control instructions.

[0220] The execution unit 307 is used to repeatedly execute the contents of the acquisition unit 303 to the fusion unit 306 to form closed-loop control until the AGV completes the tracking task of the entire fixed route.

[0221] In one embodiment, the building unit 301 includes:

[0222] The acquisition module is used to acquire the AGV's preset fixed running route data, including the path point coordinate sequence, curvature change information and slope distribution, and to build a fixed route tracking environment model.

[0223] The data acquisition and recording module is used to select oil-stained roads, water-stained roads, gravel-scattered areas, and elevation difference joint areas as typical slippery road sections in the actual operation area of ​​the factory. It collects road surface images or texture samples of each road section under different lighting conditions, and simultaneously records the wheel slip rate, heading angle fluctuation and trajectory deviation when the AGV passes through in the empty, half-loaded and full-loaded states to form actual measurement data.

[0224] A module is established to create a factory floor environment model based on measured data, which includes road surface type labels, load levels, and the mapping relationship between slippage behavior.

[0225] An integration module is used to integrate a fixed-route tracking environment model with a factory floor environment model to form an offline training environment model.

[0226] In one embodiment, the generation unit 302 includes:

[0227] The generation module is used to train and generate a fixed route tracking strategy library in the offline training environment model with the goal of minimizing the lateral deviation and heading error between the actual trajectory of the AGV and the preset route.

[0228] The training module is used in the offline training environment model to train and generate corresponding slippage compensation strategies for each road surface type and load combination, with the goal of suppressing wheel slippage, reducing attitude vibration and maintaining driving stability, in order to form a slippage compensation strategy library.

[0229] In one embodiment, the acquisition unit 303 includes:

[0230] The receiving and determining module is used to receive the fixed route task instruction currently being executed by the AGV, determine the target path point sequence and the start and end range of the task, so as to obtain task information;

[0231] The acquisition module is used to acquire real-time road surface images or texture features in front of and below the AGV using onboard cameras or vision sensors.

[0232] The acquisition module is used to acquire the position, speed, heading angle, left and right wheel speed difference and acceleration information of the AGV in real time through the inertial measurement unit, wheel speed encoder and positioning module, so as to form the motion state information of the AGV itself;

[0233] The acquisition module is used to obtain the current real-time load information of the AGV through the weighing sensor or the motor current estimation module.

[0234] In one embodiment, the identification and matching unit 304 includes:

[0235] The identification module is used to classify and identify the collected road surface images or texture features, and determine whether the current road surface belongs to oil stains, water stains, gravel, joints or normal road surface, so as to obtain the current road surface type.

[0236] The query module is used to combine the current road surface type, real-time load information and AGV current motion status information to query the slippage compensation strategy library and retrieve the initial slippage compensation strategy with the highest matching degree.

[0237] The selection module is used to choose the strategy with the closest load under the same road surface type as the initial slip compensation strategy if no perfect match is found.

[0238] In one embodiment, the acquisition fine-tuning unit 305 includes:

[0239] The configuration module is used to set the initial slippage compensation strategy as the baseline strategy for online fine-tuning;

[0240] The data acquisition module is used to continuously collect the actual motion response of the AGV, including the difference in rotational speed between the left and right drive wheels, the yaw rate of the vehicle body, and the trend of trajectory deviation.

[0241] The input module is used to input the AGV's actual motion response, road surface image or texture features, and the baseline strategy into the lightweight online reinforcement learning module.

[0242] The evaluation module is used by the online reinforcement learning module to evaluate the effectiveness of the current baseline strategy in suppressing slippage and maintaining stability based on the deviation between the current motion response and the desired driving state, so as to obtain the evaluation results.

[0243] The adjustment module is used to incrementally adjust the control parameters in the baseline strategy based on the evaluation results, and generate a dynamic slippage compensation strategy that adapts to the current instantaneous operating conditions.

[0244] In one embodiment, the fusion unit 306 includes:

[0245] The query output module is used to query and output basic control commands for maintaining path tracking from the fixed route tracking strategy library based on the current pose of the AGV and the next target path point in the preset fixed route. The basic control commands include basic linear velocity and basic angular velocity.

[0246] The extraction module is used to extract compensation control quantities that match the current working conditions from the dynamic slippage compensation strategy. The compensation control quantities include compensation linear velocity adjustment and compensation angular velocity correction for suppressing slippage.

[0247] The superposition module is used to superimpose the base linear velocity and the compensation linear velocity adjustment amount according to the preset fusion rules to obtain the target linear velocity; at the same time, it superimposes the base angular velocity and the compensation angular velocity correction amount to obtain the target angular velocity.

[0248] The integration module is used to integrate the target linear velocity and the target angular velocity to form the target AGV control command.

[0249] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned anti-skid control device 300 for fixed routes on complex road surfaces of AGV and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0250] The aforementioned anti-skid control device 300 for fixed routes on complex road surfaces used by AGVs can be implemented as a computer program, which can, for example... Figure 4 It runs on the computer device shown.

[0251] Please see Figure 4 , Figure 4 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0252] See Figure 4 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0253] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform an anti-skid control method for a fixed route on a complex road surface of an AGV.

[0254] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0255] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute an anti-skid control method for a fixed route on a complex road surface of an AGV.

[0256] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 4The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0257] The processor 502 is used to run a computer program 5032 stored in the memory to perform the following steps:

[0258] S110: Construct a fixed-route tracking environment model based on a preset fixed operating route, and construct a factory floor environment model based on measured data of slippery road surfaces in actual factory scenarios. These two models together constitute the offline training environment model. S120: In the offline training environment model, deep reinforcement learning algorithms are used to train and generate a fixed-route tracking strategy library and a slippage compensation strategy library, respectively. S130: Obtain task information for the AGV executing the fixed route, and collect in real-time road surface images or texture features at the current AGV location, AGV's own motion state information, and load information. S140: Identify the current road surface type based on the road surface image or texture features, and combine this with the A... Based on the AGV's own motion state information and load information, the corresponding initial slippage compensation strategy is matched from the slippage compensation strategy library; S150, based on the real-time collected AGV actual motion response and road surface image or texture features, the initial slippage compensation strategy is fine-tuned online to generate a dynamic slippage compensation strategy adapted to the current working conditions; S160, the basic control commands output by the fixed route tracking strategy library according to the current path point are fused with the compensation control commands output by the dynamic slippage compensation strategy to generate the target AGV control commands; S170, steps S130 to S160 are repeated to form a closed-loop control until the AGV completes the tracking task of the entire fixed route.

[0259] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0260] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0261] Therefore, the present invention also provides a storage medium. This storage medium can be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform the following steps:

[0262] S110: Construct a fixed-route tracking environment model based on a preset fixed operating route, and construct a factory floor environment model based on measured data of slippery road surfaces in actual factory scenarios. These two models together constitute the offline training environment model. S120: In the offline training environment model, deep reinforcement learning algorithms are used to train and generate a fixed-route tracking strategy library and a slippage compensation strategy library, respectively. S130: Obtain task information for the AGV executing the fixed route, and collect in real-time road surface images or texture features at the current AGV location, AGV's own motion state information, and load information. S140: Identify the current road surface type based on the road surface image or texture features, and combine this with the A... Based on the AGV's own motion state information and load information, the corresponding initial slippage compensation strategy is matched from the slippage compensation strategy library; S150, based on the real-time collected AGV actual motion response and road surface image or texture features, the initial slippage compensation strategy is fine-tuned online to generate a dynamic slippage compensation strategy adapted to the current working conditions; S160, the basic control commands output by the fixed route tracking strategy library according to the current path point are fused with the compensation control commands output by the dynamic slippage compensation strategy to generate the target AGV control commands; S170, steps S130 to S160 are repeated to form a closed-loop control until the AGV completes the tracking task of the entire fixed route.

[0263] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0264] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0265] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For example, the division of each unit is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0266] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the device of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0267] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0268] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for anti-skid control of fixed routes on complex road surfaces for AGVs, characterized in that, include: S110 constructs a fixed route tracking environment model based on a preset fixed running route, and constructs a factory floor environment model based on actual measured data of slippery road surfaces in the actual factory scenario. The two together constitute an offline training environment model. S120, in the offline training environment model, uses deep reinforcement learning algorithm to train and generate a fixed route tracking strategy library and a slippage compensation strategy library respectively. S130: Obtain task information of AGV executing a fixed route, and collect in real time the road surface image or texture features of the current AGV location, AGV's own motion status information, and load information; S140 identifies the current road surface type based on road surface images or texture features, and combines the AGV's own motion state information and load information to match the corresponding initial slippage compensation strategy from the slippage compensation strategy library; S150, based on the real-time acquisition of the AGV's actual motion response and road surface image or texture features, fine-tunes the initial slippage compensation strategy online to generate a dynamic slippage compensation strategy that adapts to the current working conditions. S160: The basic control instructions output by the fixed route tracking strategy library based on the current path point are merged with the compensation control instructions output by the dynamic slippage compensation strategy to generate the target AGV control instructions. S170, Repeat steps S130 to S160 to form closed-loop control until the AGV completes the tracking task of the entire fixed route; The method involves online fine-tuning of the initial slippage compensation strategy based on real-time acquired AGV motion response and road surface image or texture features to generate a dynamic slippage compensation strategy adapted to the current working conditions, including: The initial slippage compensation strategy is used as the baseline strategy for online fine-tuning; Continuously collect the actual motion response of the AGV, including the difference in rotational speed between the left and right drive wheels, the yaw rate of the vehicle body, and the trend of trajectory deviation; The actual motion response of the AGV, road surface image or texture features, and the baseline strategy are all input into a lightweight online reinforcement learning module. The online reinforcement learning module evaluates the effectiveness of the current baseline strategy in suppressing slippage and maintaining stability based on the deviation between the current motion response and the desired driving state, in order to obtain the evaluation results; Based on the evaluation results, the control parameters in the baseline strategy are adjusted incrementally to generate a dynamic slippage compensation strategy that adapts to the current instantaneous operating conditions.

2. The anti-skid control method for fixed routes on complex road surfaces for AGVs according to claim 1, characterized in that, The fixed route tracking environment model is constructed based on a preset fixed operating route, and the factory floor environment model is constructed based on measured data of slippery surfaces in actual factory scenarios. Together, these two constitute an offline training environment model, including: Acquire the preset fixed running route data of AGV, including the path point coordinate sequence, curvature change information and slope distribution, and construct a fixed route tracking environment model; In the actual operating area of ​​the factory, oil-stained roads, water-stained roads, gravel-scattered areas, and elevation difference joint areas were selected as typical slippery road sections. Road surface images or texture samples of each section under different lighting conditions were collected, and the wheel slip rate, heading angle fluctuation and trajectory deviation of the AGV when passing through in the empty, half-loaded and full-loaded states were recorded simultaneously to form actual measurement data. Based on measured data, a factory floor environment model was established, which includes road surface type labels, load levels and the mapping relationship between slippage behavior. The fixed route tracking environment model is integrated with the factory floor environment model to form an offline training environment model.

3. The anti-skid control method for fixed routes on complex road surfaces for AGVs according to claim 1, characterized in that, In the offline training environment model, a fixed route tracking strategy library and a slippage compensation strategy library are trained and generated using a deep reinforcement learning algorithm, including: In the offline training environment model, with the goal of minimizing the lateral deviation and heading error between the actual AGV trajectory and the preset route, a fixed route tracking strategy library is generated by training a deep reinforcement learning algorithm. In the offline training environment model, with the goals of suppressing wheel slippage, reducing attitude vibration, and maintaining driving stability, corresponding slippage compensation strategies are trained and generated for each road surface type and load combination to form a slippage compensation strategy library.

4. The anti-skid control method for fixed routes on complex road surfaces for AGVs according to claim 1, characterized in that, The process of acquiring task information for the AGV to execute a fixed route, and real-time collection of road surface images or texture features at the current location of the AGV, the AGV's own motion state information, and load information, includes: Receive the fixed route task instructions currently being executed by the AGV, determine the target path point sequence and the start and end range of the task, and obtain task information; Real-time acquisition of road surface images or texture features in front of and below the AGV using onboard cameras or vision sensors; The AGV's position, speed, heading angle, left and right wheel speed difference, and acceleration information are acquired in real time through an inertial measurement unit, wheel speed encoder, and positioning module to form the AGV's own motion state information. The real-time load information of the AGV is obtained through weighing sensors or motor current estimation modules.

5. The anti-skid control method for fixed routes on complex road surfaces for AGVs according to claim 1, characterized in that, The step of identifying the current road surface type based on road surface images or texture features, and combining this with the AGV's own motion state information and load information, to match a corresponding initial slippage compensation strategy from the slippage compensation strategy library includes: The collected road surface images or texture features are classified and identified to determine whether the current road surface belongs to oil stains, water stains, gravel, joints, or normal road surface, so as to obtain the current road surface type; By combining the current road surface type, real-time load information and the current motion status information of the AGV, the slippage compensation strategy library is queried to obtain the initial slippage compensation strategy with the highest matching degree. If no perfect match is found, the strategy with the closest load under the same road surface type is selected as the initial slip compensation strategy.

6. The anti-skid control method for fixed routes on complex road surfaces for AGVs according to claim 1, characterized in that, The process of fusing the basic control commands output by the fixed route tracking strategy library based on the current path point with the compensation control commands output by the dynamic slippage compensation strategy to generate target AGV control commands includes: Based on the current pose of the AGV and the next target path point in the preset fixed route, the basic control instructions for maintaining path tracking are queried from the fixed route tracking strategy library and output. The basic control instructions include basic linear velocity and basic angular velocity. The compensation control quantity matching the current operating condition is extracted from the dynamic slippage compensation strategy. The compensation control quantity includes the compensation linear velocity adjustment and the compensation angular velocity correction for suppressing slippage. According to the preset fusion rules, the base linear velocity and the compensation linear velocity adjustment amount are superimposed to obtain the target linear velocity; at the same time, the base angular velocity and the compensation angular velocity correction amount are superimposed to obtain the target angular velocity. The target linear velocity and target angular velocity are integrated to form the target AGV control command.

7. An anti-skid control device for fixed routes on complex road surfaces used by AGVs, characterized in that, include: The building unit is used to build a fixed route tracking environment model based on a preset fixed running route, and to build a factory floor environment model based on measured data of slippery road surfaces in actual factory scenarios. Together, the two constitute an offline training environment model. The generation unit is used to train and generate a fixed route tracking strategy library and a slippage compensation strategy library respectively in the offline training environment model using deep reinforcement learning algorithms. The acquisition and collection unit is used to acquire task information of the AGV executing a fixed route, and to collect in real time the road surface image or texture features of the current AGV location, the AGV's own motion status information, and load information; The identification and matching unit is used to identify the current road surface type based on the road surface image or texture features, and to match the corresponding initial slippage compensation strategy from the slippage compensation strategy library by combining the AGV's own motion state information and load information. The acquisition and fine-tuning unit is used to fine-tune the initial slippage compensation strategy online based on the real-time acquired AGV actual motion response and road surface image or texture features, and generate a dynamic slippage compensation strategy that adapts to the current working conditions. The fusion unit is used to fuse the basic control instructions output by the fixed route tracking strategy library based on the current path point with the compensation control instructions output by the dynamic slippage compensation strategy to generate the target AGV control instructions. The execution unit is used to repeatedly execute the content from the acquisition unit to the fusion unit to form closed-loop control until the AGV completes the tracking task of the entire fixed route; The acquisition fine-tuning unit includes: The configuration module is used to set the initial slippage compensation strategy as the baseline strategy for online fine-tuning; The data acquisition module is used to continuously collect the actual motion response of the AGV, including the difference in rotational speed between the left and right drive wheels, the yaw rate of the vehicle body, and the trend of trajectory deviation. The input module is used to input the AGV's actual motion response, road surface image or texture features, and the baseline strategy into the lightweight online reinforcement learning module. The evaluation module is used by the online reinforcement learning module to evaluate the effectiveness of the current baseline strategy in suppressing slippage and maintaining stability based on the deviation between the current motion response and the desired driving state, so as to obtain the evaluation results. The adjustment module is used to incrementally adjust the control parameters in the baseline strategy based on the evaluation results, and generate a dynamic slippage compensation strategy that adapts to the current instantaneous operating conditions.

8. A computer device, characterized in that, The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the method as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 6.

Images