Intelligent forklift axle dynamic steering control method, system and device

Abstract

The application discloses a kind of intelligent forklift axle dynamic steering control method, system and equipment, it is related to vehicle control technology relevant field, the method real-time acquisition intelligent forklift's own state, surrounding environment and task instruction data, all data are unified to the body coordinate system of intelligent forklift itself;Application double-flow neural encoder is encoded respectively to the surrounding environment point cloud data and task instruction text data of intelligent forklift, two kinds of vectors obtained by fusion coding are formed into joint representation vector;According to the joint representation vector of intelligent forklift currently, the predicted action is output through pre-training conditional variational autoencoder, and the optimal, physically feasible steering control action is sampled from the predicted action combination by combining reinforcement learning algorithm;The steering control action is kinematically feasible to check, and the final control instruction is generated by safety correction.It solves the problem that existing method responds slowly, inefficient and poor linearity, guarantees the safety of forklift operation process.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and in particular to a method, system and device for dynamic steering control of intelligent forklift axles. Background Technology

[0002] Traditional forklifts typically use fixed-axle steering, with the rear axle being the steering axle and the front axle the drive axle, offering a single-purpose mode. In contrast, intelligent forklifts employ dynamic steering control, where the drive axle and steering axle dynamically switch between multiple steering modes based on task requirements, real-time road conditions, and system commands, achieving high-precision, high-stability closed-loop control. Precise dynamic control of intelligent forklifts reduces the need for aisle width, increasing warehouse density and volume ratio; it also reduces the multiple redundant operations required by traditional forklifts to adjust their posture, making the travel path an optimal, smooth curve between two points, thus shortening the cycle time for a single operation.

[0003] Traditional control methods rely on mechanical and hydraulic power steering, which are based on open-loop control logic. This results in slow response, low efficiency, and poor speed regulation linearity, making it unable to sense and execute tasks to reach precise coordinate positions. When a forklift is fully loaded or unloaded, the vehicle's dynamic characteristics differ greatly. Traditional fixed-parameter hydraulic or electric control systems cannot automatically compensate for this change, leading to inconsistent operating feel and control effects that are difficult to predict. Furthermore, traditional methods lack data support for predictive protection and performance optimization, resulting in insufficient safety. Summary of the Invention

[0004] To address the technical problems of the prior art, this application provides an intelligent forklift axle dynamic steering control method, system, and device that can directly and adaptively map complex environments and task instructions into optimal steering actions. It applies deep neural networks to understand the point cloud data and control instruction text of the forklift operation scenario, predicts and generates the optimal steering control action, and executes it accurately in the continuous action space.

[0005] This application provides a method for dynamic steering control of an intelligent forklift axle, including:

[0006] Step S10: Establish a data acquisition system. Based on visual information acquisition equipment and various types of sensors, acquire data on the intelligent forklift's own status, surrounding environment, and task instructions. Preprocess all data and unify it into the intelligent forklift's own vehicle coordinate system. Step S20: Apply a dual-stream neural encoder to encode the point cloud data of the surrounding environment of the intelligent forklift and the text data of the task instruction, respectively, to understand the topological structure of the space in which the forklift is located, establish the association between the intention of the task to be executed and the state of the forklift, and fuse the two types of vectors obtained by encoding to form a joint representation vector. Step S30: Based on the current joint representation vector of the intelligent forklift, the predicted action is output through a pre-trained conditional variational autoencoder. Combined with a reinforcement learning algorithm, the optimal and physically feasible steering control action is sampled from the set of predicted actions. Step S40: Perform kinematic feasibility verification on the generated steering control action, and modify the action through the control barrier function to generate the final control command, ensuring that the state of the intelligent forklift is always within the preset safety set.

[0007] Furthermore, the data acquisition equipment in the data acquisition system includes: LiDAR: Acquires 3D point cloud data of the environment in which the intelligent forklift is located. Each point cloud contains three-dimensional coordinates and reflection intensity, and the sampling frequency is 10-20 frames per second. Stereo vision camera: Acquires RGB color images and depth information to identify object categories in intelligent forklift operation scenarios and supplement blind spot details; Wheel encoders and steering angle sensors: accurately measure the rotational speed of each wheel of the intelligent forklift and the actual angle of each steering axle; Central controller: Provides the current task instructions, vehicle mode, and load information read from the CAN bus of the intelligent forklift; High-precision positioning system: The positioning module integrates visual SLAM to provide the position and direction of travel of the intelligent forklift in the global map of the working environment.

[0008] Deploy a precise time protocol for all data acquisition devices to ensure that the start time of a LiDAR scan, the moment of camera exposure, and the moment of sensor sampling have a unified, microsecond-level precision timestamp; transform all data collected by all devices into a vehicle coordinate system with the intelligent forklift's center of mass as the origin and the direction of the forklift's head as the positive X-axis, based on the calibration matrix of their installation position and angle; represent the point cloud obtained by LiDAR scanning, the obstacle box identified by the camera, and the position data collected by the positioning system as (x, y, z) three-dimensional coordinates related to the vehicle body.

[0009] Furthermore, the dual-stream neural encoder consists of an environment-stream encoder and a task-state-stream encoder, extracting feature information from preprocessed 3D point cloud data and task instruction text data, and mapping them into mathematical vectors that are easy for computers to process: The environmental flow encoder uses a backbone network to extract point cloud features of the environment in which the intelligent forklift is located. First, it uses farthest point sampling to filter out key points from the environmental point cloud: Randomly select a point from the preprocessed 3D point cloud data as the initial point and add it to the sampling point set. Calculate the Euclidean distance between the initial point and all other points, and select the point with the farthest distance to add to the sampling point set. Continue to select the data with the farthest distance from the sampling points in the set as the new sampling point from other points until the preset number of sampling points is met. Secondly, for each sampling keypoint, K neighboring points are selected from its local region to form a local neighborhood. The neighborhood points are searched from a spherical region with radius R centered at the sampling point. Convolution is performed in the local neighborhood of each sampling point to learn local geometric features, which are used to determine whether each point constitutes a plane, an edge, or a corner. Then, multiple multilayer perceptron structures are applied to expand the receptive field area through linear processing and neuron activation, extract high-level feature representations, and obtain a deeper understanding of environmental information. After multi-layer perception, the network finally aggregates the features of all point clouds into a fixed-length global environment feature vector through global max pooling. Based on real-time acquired RGB color images and depth information, the vector representation is optimized to implicitly encode the shape and size of passable areas, the distribution, type and dynamics of obstacles, and the topology of the intelligent forklift operation scenario, including passages, intersections and dead ends.

[0010] The task-state flow encoder takes task instructions, vehicle status, and historical action sequences as input. The task instructions are processed into natural language text data. The vehicle status includes the current speed, heading angle, steering angle of each axle, load, and tilt angle of the intelligent forklift. The historical action sequence records the sequence of control instructions issued by the controller over a period of time, including the target steering angle of the front axle, the target steering angle of the rear axle, and the target speed. The encoder first processes the text data through the Transformer, converting the task instructions into task semantic vectors, encoding the operation type, target location, behavioral constraints, and priority information involved in the instructions. For interrupt instructions, a high-priority flag is set; for compound tasks, a flag indicating the order of subtasks is set. For vehicle state data, the encoder applies a one-dimensional temporal convolutional network in parallel to learn the state characteristics of the intelligent forklift at the current moment, and captures complex behavioral patterns by deepening the network layers to obtain a state sequence feature vector representing the dynamic characteristics of the vehicle. In addition, the historical action sequence is input into the recurrent neural network to understand the behavioral patterns of the intelligent forklift in the process of executing the most recent task, reflecting the recent policy tendency of the vehicle controller, and obtaining a historical action feature vector. The task semantic vector, state sequence feature vector, and historical action feature vector are concatenated. The task semantic vector is used as the query, and cross-attention is used to modulate the concatenated vector to enhance the feature elements in the state features that are related to the task semantics. The modulated vector is then projected through a linear layer to output the task-state context vector.

[0011] Furthermore, environmental, task, vehicle status, and action execution data from the historical work records of the intelligent forklift are collected. These data are then processed by a dual-stream neural encoder to obtain historical joint vectors, with the corresponding control commands serving as the true labels. Data augmentation techniques are used to expand the intelligent forklift's operating scenarios, increasing the proportion of data in narrow passages, sharp turns, and vehicle docking scenarios. These data are combined into a dataset for training a conditional variational autoencoder, resulting in a pre-trained model. This model processes the current joint vector of the intelligent forklift in real time to predict the vehicle's steering actions. A reward function for the reinforcement learning algorithm is set, considering task completion, efficiency, and safety. A delayed target network update strategy is applied to complete the reinforcement learning model training, resulting in a pre-trained Q-network used for action selection. In the practical application of intelligent forklift dynamic control, the joint representation vector perceived by each forklift at the current moment is obtained in real time from the dual-stream neural encoder, and input into the pre-trained conditional variational autoencoder to generate N candidate actions. Each candidate action and the current joint representation vector are input into the pre-trained Q network to calculate the Q value of each action. The candidate actions are sorted in descending order according to the Q value, and the candidate action with the highest Q value is selected as the optimal solution to generate the final forklift axle steering control command.

[0012] Furthermore, the detailed steps for verifying kinematic feasibility include: For the optimal steering control action selected by reinforcement learning, the operational model is used to perform trajectory forward simulation, predict the position, speed, yaw rate and sideslip angle of the intelligent forklift in the next few time steps, generate the predicted trajectory, and check whether it is within the safe range. Calculate the instantaneous turning radius of the forklift under the steering angle set in the optimal control steering action, and compare it with the theoretical minimum turning radius to check whether the instantaneous radius is greater than the minimum radius; Calculate the rate of change required from the current steering angle to the target steering angle, and ensure that the rate of change does not exceed the maximum allowable speed of the steering servo motor; otherwise, the command cannot be executed in time. Analyze the rate of change of curvature of the predicted trajectory, i.e. the abruptness of the steering, and check whether the rate of change of curvature is lower than the preset threshold to ensure smooth steering control. If a candidate action satisfies all the constraints of the feasibility check, it is marked as "kinematically feasible" and proceeds to the next step of safety correction. If the turning radius is slightly less than the minimum value, the constraint weighted adjustment algorithm is activated. While maintaining the original turning trend as much as possible, the turning angle is adjusted with the minimum magnitude along the gradient direction of the constraint boundary to satisfy the constraints. If the turning angular velocity far exceeds the actuator's capability, the candidate action is directly rejected, the reason is recorded, and fed back to the reinforcement learning model as a negative sample for learning. The safety correction process defines two safety sets: a collision avoidance safety set and a rollover avoidance safety set, to ensure that the distance between the intelligent forklift and the nearest obstacle, as well as the rollover angle of the vehicle body, are within a safe range; a control barrier function is set for each safety set, and if the function value is greater than or equal to zero, it means that the currently generated steering control command meets the safety standards; The current vehicle status, steering control actions, operational model, and control barrier function of the intelligent forklift are input into the real-time optimization solver. The solver quickly calculates the optimal safe action that satisfies all safety constraints, corrects the currently set steering control actions, adjusts the control values ​​of the front axle steering angle, rear axle steering angle, and speed, and generates the final control command to control the intelligent forklift to complete dynamic steering.

[0013] This application also provides an intelligent forklift axle dynamic steering control system, including: Data acquisition and processing module: Used to establish a data acquisition system, based on visual information acquisition equipment and various types of sensors, to acquire data on the intelligent forklift's own status, surrounding environment and task instructions, and to preprocess all data and unify it into the intelligent forklift's own vehicle coordinate system; Dual-stream neural coding module: The dual-stream neural encoder is used to encode the point cloud data of the surrounding environment of the intelligent forklift and the text data of the task instruction, respectively. It understands the topological structure of the space in which the forklift is located, establishes the association between the intention of the task to be executed and the state of the forklift, and fuses the two types of vectors obtained by encoding to form a joint representation vector. Steering action decision module: Based on the current joint representation vector of the intelligent forklift, it outputs the predicted action through a pre-trained conditional variational autoencoder, and combines it with a reinforcement learning algorithm to sample the optimal and physically feasible steering control action from the set of predicted actions. Control command verification module: This module verifies the kinematic feasibility of the generated steering control actions and corrects the actions using a control barrier function to generate the final control commands, ensuring that the intelligent forklift remains within a preset safety set.

[0014] This application also proposes an intelligent forklift axle dynamic steering control device, the device comprising: a memory, a processor, and programs such as an intelligent forklift axle dynamic steering control algorithm stored in the memory and executable on the processor, wherein the intelligent forklift axle dynamic steering control algorithm and other programs are steps for implementing the intelligent forklift axle dynamic steering control method described above.

[0015] This application also provides a computer program product, which includes programs such as an intelligent forklift axle dynamic steering control algorithm. When the intelligent forklift axle dynamic steering control algorithm and other programs are executed by a processor, they implement an intelligent forklift axle dynamic steering control method as described above.

[0016] This application discloses the following technical effects: This application provides a method, system, and device for dynamic steering control of intelligent forklift axles, enabling intelligent forklifts to make smoother, more efficient, and safer maneuvering decisions in complex, dynamic, and non-standard real warehouse environments, thereby achieving a qualitative improvement in the overall efficiency and flexibility of the logistics handling system. Specifically, the method proposed in this application abandons the concept of discrete patterns and uses neural networks to transform steering control into a search problem in a high-dimensional continuous space. It searches for the optimal solution for the current operating scenario in an infinite continuous space composed of any combination of front and rear wheel steering angles, which is more in line with actual geometric constraints and reduces the decision error rate. This method learns the scenario topology and understands the task content through a dual-stream neural encoder, providing a deep, task-related spatial understanding, which is the basis for making forward-looking decisions. Furthermore, based on an end-to-end learning process, this method applies a pre-trained autoencoder to map feature vectors to control actions, and learns the forklift response under various load, tire pressure, and ground friction conditions during training. This implicitly compensates for errors in the dynamic model and unmodeled dynamics, demonstrating strong robustness. After obtaining the steering control action, this method evaluates the reliability of the decision based on reinforcement learning and safety checks, improving the safety of intelligent forklift operation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly described below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from these processes.

[0018] Figure 1 This is a flowchart illustrating the intelligent forklift axle dynamic steering control method provided in an embodiment of this application.

[0019] Figure 2 This is a schematic diagram of the structure of the intelligent forklift axle dynamic steering control system provided in an embodiment of this application. Detailed Implementation

[0020] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description of this application will be provided in conjunction with the accompanying drawings. The described embodiments should not be considered as limitations on this application. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] In the following description, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used herein is for the purpose of describing embodiments of this application only.

[0023] Example 1: This application provides a method for dynamic steering control of an intelligent forklift axle, such as... Figure 1 As shown, the method includes: Step S10: Establish a data acquisition system. Based on visual information acquisition equipment and various types of sensors, acquire data on the intelligent forklift's own status, surrounding environment, and task instructions. Preprocess all data and unify it into the intelligent forklift's own vehicle coordinate system.

[0024] In this embodiment, the data acquisition device in the data acquisition system includes: LiDAR: Acquires 3D point cloud data of the environment in which the intelligent forklift is located. Each point cloud contains three-dimensional coordinates and reflection intensity, and the sampling frequency is 10-20 frames per second. Stereo vision camera: Acquires RGB color images and depth information to identify object categories in intelligent forklift operation scenarios and supplement blind spot details; Wheel encoders and steering angle sensors: accurately measure the rotational speed of each wheel of the intelligent forklift and the actual angle of each steering axle; Central controller: Provides the current task instructions, vehicle mode, and load information read from the CAN bus of the intelligent forklift; High-precision positioning system: A positioning module integrating visual SLAM provides the position and direction of travel of the intelligent forklift in a global map of the working environment; Deploy a precise time protocol for all data acquisition devices to ensure that the start time of a LiDAR scan, the moment of camera exposure, and the moment of sensor sampling have a unified timestamp with microsecond-level precision. All data collected by the devices are uniformly transformed into a vehicle coordinate system with the center of mass of the intelligent forklift as the origin and the direction of the forklift head as the positive X-axis, based on the calibration matrix of their installation position and angle. The point cloud obtained by the LiDAR scan, the obstacle box identified by the camera, and the position data collected by the positioning system are all represented as (x, y, z) three-dimensional coordinates related to the vehicle body. The acquired 3D point cloud data is preprocessed to facilitate subsequent encoder modeling, constructing a 3D scene of the environment in which the intelligent forklift operates, and obtaining spatial topology information. The steps include point cloud filtering, elevation normalization, subdividing the land into subplots, point cloud denoising, downsampling, and coordinate normalization. First, a cloth simulation algorithm is used to filter the point cloud, dividing the point cloud data into ground points and non-ground points. Then, CloudCompare software is used to normalize the elevation of the point cloud. Secondly, the operation scenario of the intelligent forklift is divided into multiple sub-regions to reduce the complexity of the data processed by the environmental flow encoder. The local density of the point cloud in each sub-region is calculated using statistical methods. Neighborhood radius and threshold are set. Points with fewer points in the neighborhood than the threshold are marked as noise points and removed. Finally, in the downsampling step, the region plane is divided into multiple grids in the Cartesian coordinate system, and the highest point in each grid is taken as the sampling point. The coordinates of the point cloud in each region are normalized, and the point cloud coordinate values ​​are unified to the range of [0,1].

[0025] Step S20: Apply a dual-stream neural encoder to encode the point cloud data of the surrounding environment of the intelligent forklift and the text data of the task instruction, respectively, to understand the topological structure of the space in which the forklift is located, establish the association between the intention of the task to be executed and the state of the forklift, and fuse the two types of vectors obtained by encoding to form a joint representation vector.

[0026] In this embodiment, the dual-stream neural encoder consists of an ambient stream encoder and a task-state stream encoder. It extracts feature information from the preprocessed 3D point cloud data and task instruction text data, and maps them into mathematical vectors that are easy for computers to process. The environmental flow encoder uses a backbone network to extract point cloud features of the environment in which the intelligent forklift is located. First, it uses farthest point sampling to filter out key points from the environmental point cloud: Randomly select a point from the preprocessed 3D point cloud data as the initial point and add it to the sampling point set. Calculate the Euclidean distance between the initial point and all other points, and select the point with the farthest distance to add to the sampling point set. Continue to select the data with the farthest distance from the sampling points in the set as the new sampling point from other points until the preset number of sampling points is met. Secondly, for each sampling keypoint, K neighboring points are selected from its local region to form a local neighborhood. The neighborhood points are searched from a spherical region with radius R centered at the sampling point. Convolution is performed in the local neighborhood of each sampling point to learn local geometric features, which are used to determine whether each point constitutes a plane, an edge, or a corner. Then, multiple multilayer perceptron structures are applied to expand the receptive field area through linear processing and neuron activation, extract high-level feature representations, and obtain a deeper understanding of environmental information. After multi-layer perception, the network finally aggregates the features of all point clouds into a fixed-length global environment feature vector through global max pooling. Based on real-time acquired RGB color images and depth information, the vector representation is optimized to implicitly encode the shape and size of passable areas, the distribution, type and dynamics of obstacles, and the topology of the intelligent forklift operation scenario, including passages, intersections and dead ends.

[0027] The task-state flow encoder takes task instructions, vehicle status, and historical action sequences as input. The task instructions are processed into natural language text data. The vehicle status includes the current speed, heading angle, steering angle of each axle, load, and tilt angle of the intelligent forklift. The historical action sequence records the sequence of control instructions issued by the controller over a period of time, including the target steering angle of the front axle, the target steering angle of the rear axle, and the target speed. The encoder first processes the text data through the Transformer, converting the task instructions into task semantic vectors, encoding the operation type, target location, behavioral constraints, and priority information involved in the instructions. For interrupt instructions, a high-priority flag is set; for compound tasks, a flag indicating the order of subtasks is set. For vehicle state data, the encoder applies a one-dimensional temporal convolutional network in parallel to learn the state characteristics of the intelligent forklift at the current moment, and captures complex behavioral patterns by deepening the network layers to obtain a state sequence feature vector representing the dynamic characteristics of the vehicle. In addition, the historical action sequence is input into the recurrent neural network to understand the behavioral patterns of the intelligent forklift in the process of executing the most recent task, reflecting the recent policy tendency of the vehicle controller, and obtaining a historical action feature vector. The task semantic vector, state sequence feature vector and historical action feature vector are concatenated. The task semantic vector is used as the query. Cross attention is used to modulate the concatenated vector to enhance the feature elements in the state features that are related to the task semantics. The modulated vector is then projected through a linear layer to output the task-state context vector. Finally, the dual-stream neural encoder outputs a global environmental feature vector and a task-state context vector. Through a cross-attention mechanism, using the task-state context vector as a query, it expands the global environmental feature vector as a key and value to generate a task-modulated joint representation vector. This vector contains a focused understanding of the current task, essentially allowing the task flow to actively ask questions of the environmental flow. For example, when the task vector carries the intention of "picking up goods", cross-attention calculation is used to give higher weights to the area near the shelf, the protruding edge of the pallet, and the docking path in the environmental features; conversely, if the task is "emergency obstacle avoidance", the focus is on the direction of movement of dynamic obstacles and the nearest escape route.

[0028] Step S30: Based on the current joint representation vector of the intelligent forklift, the predicted action is output through a pre-trained conditional variational autoencoder. Combined with a reinforcement learning algorithm, the optimal and physically feasible steering control action is sampled from the set of predicted actions.

[0029] In this embodiment, environmental, task, vehicle status, and action execution data from the historical work records of the intelligent forklift are collected. These data are then processed by a dual-stream neural encoder to obtain historical joint vectors. The corresponding control commands are used as the real labels. Data augmentation techniques are employed to expand the intelligent forklift's operating scenarios, increasing the proportion of data in narrow passages, sharp turns, and vehicle docking scenarios. These data are combined into a dataset for training a conditional variational autoencoder, resulting in a pre-trained model. This model processes the current joint vector of the intelligent forklift in real time to predict the vehicle's steering actions. A reward function for the reinforcement learning algorithm is set, considering task completion, efficiency, and safety. A delayed target network update strategy is applied to complete the reinforcement learning model training, resulting in a pre-trained Q-network used for action selection. In the practical application of intelligent forklift dynamic control, the joint representation vector perceived by each forklift at the current moment is obtained in real time from the dual-stream neural encoder, and input into the pre-trained conditional variational autoencoder to generate N candidate actions. Each candidate action and the current joint representation vector are input into the pre-trained Q network to calculate the Q value of each action. The candidate actions are sorted in descending order according to the Q value, and the candidate action with the highest Q value is selected as the optimal solution to generate the final forklift axle steering control command.

[0030] Step S40: Perform kinematic feasibility verification on the generated steering control action, and modify the action through the control barrier function to generate the final control command, ensuring that the state of the intelligent forklift is always within the preset safety set.

[0031] In this embodiment, the left and right wheels of the intelligent forklift axle are merged into a single virtual wheel located at the center of the axle, and a dual-axle steering nonlinear model is established to... Let be the state vector of the forklift. To control the input vector, the model's state equation is expressed as:

[0032] in, and These represent the current and updated x-coordinate positions of the forklift's center of mass in the coordinate system, respectively. and These represent the current and updated ordinate positions of the forklift's center of mass in the coordinate system, respectively. and These represent the current and updated yaw rates of the forklift, respectively. and The current and updated speeds of the forklift. Indicates the longitudinal acceleration of the forklift. and These represent the steering angles of the forklift's front and rear axles, respectively. This indicates the forklift's sideslip angle. This indicates the wheelbase between the front and rear axles of the forklift. and These represent the distances from the front axle and rear axle to the center of mass, respectively. The detailed steps for kinematic feasibility verification include: For the optimal steering control action selected by reinforcement learning, the above-mentioned operational model is used to perform trajectory forward simulation, predict the position, speed, yaw rate and sideslip angle of the intelligent forklift in the next few time steps, generate the predicted trajectory, and check whether it is within the safe range. Calculate the instantaneous turning radius of the forklift under the steering angle set in the optimal control steering action, and compare it with the theoretical minimum turning radius to check whether the instantaneous radius is greater than the minimum radius; Calculate the rate of change required from the current steering angle to the target steering angle, and ensure that the rate of change does not exceed the maximum allowable speed of the steering servo motor; otherwise, the command cannot be executed in time. Analyze the rate of change of curvature of the predicted trajectory, i.e. the abruptness of the steering, and check whether the rate of change of curvature is lower than the preset threshold to ensure smooth steering control.

[0033] If a candidate action satisfies all the constraints of the feasibility check, it is marked as "kinematically feasible" and proceeds to the next step of safety correction. If the turning radius is slightly less than the minimum value, the constraint weighted adjustment algorithm is activated. While maintaining the original turning trend as much as possible, the turning angle is adjusted with the minimum magnitude along the gradient direction of the constraint boundary to satisfy the constraints. If the turning angular velocity far exceeds the actuator's capability, the candidate action is directly rejected, the reason is recorded, and fed back to the reinforcement learning model as a negative sample for learning.

[0034] The safety correction process defines two safety sets: a collision avoidance safety set and a rollover avoidance safety set, to ensure that the distance between the intelligent forklift and the nearest obstacle, as well as the rollover angle of the vehicle body, are within a safe range; a control barrier function is set for each safety set, and if the function value is greater than or equal to zero, it means that the currently generated steering control command meets the safety standards; The current vehicle status, steering control actions, operational model, and control barrier function of the intelligent forklift are input into the real-time optimization solver. The solver quickly calculates the optimal safe action that satisfies all safety constraints, corrects the currently set steering control actions, adjusts the control values ​​of the front axle steering angle, rear axle steering angle, and speed, and generates the final control command to control the intelligent forklift to complete dynamic steering.

[0035] Example 2: The intelligent forklift axle dynamic steering control system provided in this embodiment of the invention can execute the intelligent forklift axle dynamic steering control method provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects of the execution method, such as... Figure 2 As shown, it includes the following modules: Data acquisition and processing module: Used to establish a data acquisition system, based on visual information acquisition equipment and various types of sensors, to acquire data on the intelligent forklift's own status, surrounding environment and task instructions, and to preprocess all data and unify it into the intelligent forklift's own vehicle coordinate system; Dual-stream neural coding module: The dual-stream neural encoder is used to encode the point cloud data of the surrounding environment of the intelligent forklift and the text data of the task instruction, respectively. It understands the topological structure of the space in which the forklift is located, establishes the association between the intention of the task to be executed and the state of the forklift, and fuses the two types of vectors obtained by encoding to form a joint representation vector. Steering action decision module: Based on the current joint representation vector of the intelligent forklift, it outputs the predicted action through a pre-trained conditional variational autoencoder, and combines it with a reinforcement learning algorithm to sample the optimal and physically feasible steering control action from the set of predicted actions. Control command verification module: This module verifies the kinematic feasibility of the generated steering control actions and corrects the actions using a control barrier function to generate the final control commands, ensuring that the intelligent forklift remains within a preset safety set.

[0036] Example 3: This application provides an intelligent forklift axle dynamic steering control device, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute an intelligent forklift axle dynamic steering control method as described in the above embodiment.

[0037] In embodiment four, this application provides a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network via a communication system, or installed from a storage system. When the computer program is executed by a processing system, it performs the functions defined in the method of embodiment one of this application.

[0038] Although this application makes various references to certain modules in the system according to the embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy distinction between each other and are not used to limit the scope of protection of this invention.

[0039] The specific embodiments described above do not constitute a limitation on the scope of protection of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application. In some cases, the actions or steps described in this application can be performed in a different order than that shown in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A dynamic steering control method for intelligent forklift axles, characterized in that, The method includes: Step S10: Establish a data acquisition system. Based on visual information acquisition equipment and various types of sensors, acquire data on the intelligent forklift's own status, surrounding environment, and task instructions. Preprocess all data and unify it into the intelligent forklift's own vehicle coordinate system. Step S20: Apply a dual-stream neural encoder to encode the point cloud data of the surrounding environment of the intelligent forklift and the text data of the task instructions, respectively, to understand the topological structure of the space in which the forklift is located, establish the association between the intention of the task to be executed and the state of the forklift, and fuse the two types of vectors obtained by encoding to form a joint representation vector: The dual-stream neural encoder consists of an environment flow encoder and a task-state flow encoder. It extracts feature information from preprocessed 3D point cloud data and task instruction text data, maps them into mathematical vectors that are easy for computers to process, and outputs a global environment feature vector and a task-state context vector to form a joint representation vector. The task-state flow encoder takes task instructions, vehicle status, and historical action sequences as inputs. The task instructions are processed into natural language text data. The vehicle status includes the current speed, heading angle, steering angle of each axle, load, and tilt angle of the intelligent forklift. The historical action sequence records the sequence of control instructions issued by the controller over a period of time, including the target steering angle of the front axle, the target steering angle of the rear axle, and the target speed. The encoder processes text data through the Transformer, converting task instructions into task semantic vectors, encoding the operation type, target location, behavioral constraints, and priority information involved in the instructions; For vehicle state data, the encoder applies a one-dimensional temporal convolutional network in parallel to learn the state characteristics of the intelligent forklift at the current moment, and captures complex behavioral patterns by deepening the network layers to obtain a state sequence feature vector representing the dynamic characteristics of the vehicle; the historical action sequence is input into the recurrent neural network to understand the behavioral patterns of the intelligent forklift in the process of executing the most recent task, reflecting the recent strategy tendency of the vehicle controller, and obtaining the historical action feature vector. The task semantic vector, state sequence feature vector and historical action feature vector are concatenated. The task semantic vector is used as the query. Cross attention is used to modulate the concatenated vector to enhance the feature elements in the state features that are related to the task semantics. The modulated vector is then projected through a linear layer to output the task-state context vector. Step S30: Based on the current joint representation vector of the intelligent forklift, the predicted action is output through a pre-trained conditional variational autoencoder. Combined with a reinforcement learning algorithm, the optimal and physically feasible steering control action is sampled from the set of predicted actions. Step S40: Perform kinematic feasibility verification on the generated steering control action, and modify the action through the control barrier function to generate the final control command, ensuring that the state of the intelligent forklift is always within the preset safety set.

2. The intelligent forklift axle dynamic steering control method as described in claim 1, characterized in that, In step S10, the data acquisition devices in the data acquisition system include: LiDAR: Acquires 3D point cloud data of the environment in which the intelligent forklift is located. Each point cloud contains three-dimensional coordinates and reflection intensity, and the sampling frequency is 10-20 frames per second. Stereo vision camera: Acquires RGB color images and depth information to identify object categories in intelligent forklift operation scenarios and supplement blind spot details; Wheel encoders and steering angle sensors: accurately measure the rotational speed of each wheel of the intelligent forklift and the actual angle of each steering axle; Central controller: Provides the current task instructions, vehicle mode, and load information read from the CAN bus of the intelligent forklift; High-precision positioning system: The positioning module integrates visual SLAM to provide the position and direction of travel of the intelligent forklift in the global map of the working environment.

3. The intelligent forklift axle dynamic steering control method as described in claim 1, characterized in that, The environmental flow encoder uses a backbone network to extract point cloud features of the environment in which the intelligent forklift is located: Key points are selected from the environmental point cloud by sampling from the farthest point. For each sampled key point, K neighboring points are selected from its local region to form a local neighborhood. Convolution is performed in the local neighborhood of each sampled point to learn local geometric features, which are used to determine whether each point forms a plane, an edge, or a corner. Multiple multilayer perceptron structures are applied to expand the receptive field area through linear processing and neuron activation, extract high-level feature representations, and obtain a deep understanding of environmental information. After multi-layer perception, the network finally aggregates the features of all point clouds into a fixed-length global environment feature vector through global max pooling. Based on real-time acquired RGB color images and depth information, the vector representation is optimized to implicitly encode the shape and size of passable areas, the distribution, type and dynamics of obstacles, and the topology of the intelligent forklift operation scenario.

4. The intelligent forklift axle dynamic steering control method as described in claim 1, characterized in that, In step S30, environmental, task, vehicle status and action execution data from the historical work records of the intelligent forklift are collected. The historical joint vector is obtained through a dual-stream neural encoder. The corresponding control command is used as the real label. The operation scenario of the intelligent forklift is expanded through data augmentation technology and combined into a dataset for training a conditional variational autoencoder to obtain a pre-trained model. The model is then used to process the current joint vector of the intelligent forklift in real time to predict the vehicle's steering action. The reward function of the reinforcement learning algorithm is set, taking into account task completion, efficiency and safety. After delaying the target network update strategy, the reinforcement learning model is trained to obtain a pre-trained Q network for action selection.

5. The intelligent forklift axle dynamic steering control method as described in claim 1, characterized in that, Step S40, the detailed steps for kinematic feasibility verification, include: For the optimal steering control action selected by reinforcement learning, the operational model is used to perform trajectory forward simulation, predict the position, speed, yaw rate and sideslip angle of the intelligent forklift in the next few time steps, generate the predicted trajectory, and check whether it is within the safe range. Calculate the instantaneous turning radius of the forklift under the steering angle set in the optimal control steering action, and compare it with the theoretical minimum turning radius to check whether the instantaneous radius is greater than the minimum radius; Calculate the rate of change required from the current steering angle to the target steering angle, and ensure that the rate of change does not exceed the maximum allowable speed of the steering servo motor; otherwise, the command cannot be executed in time. Analyze the rate of change of curvature of the predicted trajectory, i.e. the abruptness of the steering, and check whether the rate of change of curvature is lower than the preset threshold to ensure smooth steering control.

6. The intelligent forklift axle dynamic steering control method as described in claim 1, characterized in that, In step S40, the safety correction step defines two safety sets, including a collision avoidance safety set and a rollover avoidance safety set, to ensure that the distance between the intelligent forklift and the nearest obstacle, as well as the rollover angle of the vehicle body, are within a safe range; a control barrier function is set for each safety set, and if the function value is greater than or equal to zero, it means that the currently generated steering control command meets the safety standard; The current vehicle status, steering control actions, operational model, and control barrier function of the intelligent forklift are input into the real-time optimization solver. The solver quickly calculates the optimal safe action that satisfies all safety constraints, corrects the currently set steering control actions, adjusts the control values ​​of the front axle steering angle, rear axle steering angle, and speed, and generates the final control command to control the intelligent forklift to complete dynamic steering.

7. An intelligent forklift axle dynamic steering control system, characterized in that, The system is used to implement the intelligent forklift axle dynamic steering control method according to any one of claims 1-6, the system comprising: Data acquisition and processing module: Used to establish a data acquisition system, based on visual information acquisition equipment and various types of sensors, to acquire data on the intelligent forklift's own status, surrounding environment and task instructions, and to preprocess all data and unify it into the intelligent forklift's own vehicle coordinate system; Dual-stream neural coding module: The dual-stream neural encoder is used to encode the point cloud data of the surrounding environment of the intelligent forklift and the text data of the task instruction, respectively. It understands the topological structure of the space in which the forklift is located, establishes the association between the intention of the task to be executed and the state of the forklift, and fuses the two types of vectors obtained by encoding to form a joint representation vector. Steering action decision module: Based on the current joint representation vector of the intelligent forklift, it outputs the predicted action through a pre-trained conditional variational autoencoder, and combines it with a reinforcement learning algorithm to sample the optimal and physically feasible steering control action from the set of predicted actions. Control command verification module: This module verifies the kinematic feasibility of the generated steering control actions and corrects the actions using a control barrier function to generate the final control commands, ensuring that the intelligent forklift remains within a preset safety set.

8. A dynamic steering control device for intelligent forklift axles, characterized in that, The intelligent forklift axle dynamic steering control device includes: The system includes a memory, a processor, and an intelligent forklift axle dynamic steering control program stored in the memory and executable on the processor, wherein the intelligent forklift axle dynamic steering control program, when executed by the processor, implements the intelligent forklift axle dynamic steering control method as described in any one of claims 1 to 6.

Images