Unmanned agricultural machine autonomous departure safety navigation method and system based on end-to-end model learning

Abstract

The invention discloses an end-to-end model learning-based safe navigation method and system for autonomous departure of an unmanned agricultural machine. The method comprises the following steps of: establishing an optimized objective function of an unmanned agricultural machine departure navigation task; constructing a neural network encoder based on model expansion, taking point cloud information of the laser radar as input, extracting potential distance features, and realizing interpretable mapping from point cloud data to geometric features; based on the potential distance characteristics, a control obstacle function is utilized to construct a safety constraint, and it is ensured that the unmanned agricultural machine always keeps a safety distance from a static obstacle and a dynamic obstacle in the warehouse leaving process; a learnable model prediction motion controller is designed, a dynamic and feasible local trajectory is generated in combination with a system kinematics model, distance features and safety constraints, and autonomous depot and safety control in a complex indoor environment are realized. According to the method, the autonomous navigation capability of the unmanned agricultural machine in a complex indoor environment is remarkably enhanced, and the influence of external disturbance and environment uncertainty on the control performance is effectively inhibited.

Description

Technical Field

[0001] This invention belongs to the field of autonomous navigation technology for agricultural machinery, and relates to a method and system for safe autonomous departure navigation of unmanned agricultural machinery based on end-to-end model learning. Background Technology

[0002] With increasing global food demand and a continued decline in the rural labor force, smart agriculture has become an inevitable trend in agricultural modernization. Unmanned agricultural machinery, as an important component of smart agriculture, is widely used in sowing, plant protection, tillage, and harvesting, offering new avenues for agricultural production through its unmanned and efficient operation capabilities. In most actual agricultural machinery operation scenarios, the typical workflow for unmanned agricultural machinery is "indoor parking - departure from the warehouse - arrival at the work area - execution of the work task." However, existing research and engineering implementations mainly focus on autonomous navigation and path tracking in outdoor scenarios, heavily relying on satellite positioning, high-precision maps, or visual information in open environments. In contrast, the autonomous departure from the warehouse from indoors to outdoors, as the starting point of the entire operation chain, has long lacked systematic research and effective solutions. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method and system for autonomous departure safety navigation of unmanned agricultural machinery based on end-to-end model learning. Applicable to various agricultural equipment, it effectively improves the accuracy, safety, and robustness of unmanned agricultural machinery in autonomous departure from complex indoor environments by constructing an integrated perception and control objective function, a model-based neural network encoder, control obstacle function constraints, and a learnable motion controller that predicts motion from the model. This contributes to the high-quality development of smart agriculture.

[0004] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0005] A method for safe autonomous departure navigation of unmanned agricultural machinery based on end-to-end model learning:

[0006] Based on the optimization goal of end-to-end integrated perception and control, an objective function is established for the indoor autonomous and safe navigation of unmanned agricultural machinery.

[0007] Point cloud data acquired by LiDAR is used to build a model-based neural network encoder to extract latent distance features from the point cloud data. ;

[0008] Using the potential distance features as input, construct safety constraints for the control barrier function;

[0009] By combining the objective function, potential distance features, and safety constraints, a learnable model predictive motion controller is designed. Based on the current state and point cloud features of the unmanned agricultural machine, the controller predicts the future trajectory of the unmanned agricultural machine and generates control commands, which the unmanned agricultural machine then executes.

[0010] Furthermore, the objective function is:

[0011]

[0012] Where H represents the total control duration; This indicates the attitude information of the unmanned agricultural machinery at various times. Agricultural machinery posture at all times , Represents coordinates in a local coordinate system. Indicates the heading angle of the unmanned agricultural machinery; This represents the control vector of the unmanned agricultural machinery at each moment. They represent The speed and front wheel angle of the unmanned agricultural machinery at all times; The feasible region for state and control inputs. and For adaptive weighting coefficients, express Reference attitude of unmanned agricultural machinery at +1 moment express Reference control parameters for unmanned agricultural machinery at all times.

[0013] Furthermore, Real-time attitude information of unmanned agricultural machinery ,in, , and It is the gain matrix, and , , ,in, The wheelbase represents the length of the unmanned agricultural machinery. For control vectors, To indicate Speed ​​reference control variable for unmanned agricultural machinery at all times To indicate Reference control value for the front wheel steering angle of unmanned agricultural machinery at all times.

[0014] Furthermore, , ,in, and All of these are initial weighting coefficients. , All are adjustment coefficients. The normalized distance between unmanned agricultural machinery and obstacles. For minimum safe distance, for The actual minimum distance between the shape of the unmanned agricultural machinery and the point cloud at any given moment.

[0015] Furthermore, ,in, for Real-time lidar point cloud aggregation, Indicates the attitude of agricultural machinery The shape of the vehicle body below, It is the normal vector of each face of the unmanned agricultural machinery. These are the coordinates of a point on the vehicle surface in the vehicle coordinate system. Represents the intercept of the corresponding vehicle body surface, rotation matrix ,and .

[0016] Furthermore, the potential distance feature is:

[0017]

[0018] in, , All are optimizations and iterations. The subsequent dual variable is calculated as follows:

[0019]

[0020] in, Indicates the learning rate. For projection operators, Let Lagrange's dual function be the augmented Lagrange distance from the point cloud to the unmanned agricultural machinery. This represents the gradient symbol with λ as the independent variable. This represents the gradient symbol with μ as the independent variable.

[0021] Furthermore, the safety constraint for the control barrier function is:

[0022]

[0023] in, , As slack variables, The control barrier function has the following form:

[0024]

[0025] in, It is a moment An approximate estimate of the minimum distance from the vehicle body to the point cloud, obtained by unmanned agricultural machinery using latent distance features, is calculated using the following formula:

[0026]

[0027] st .

[0028] Furthermore, the learnable model predicts the motion controller as follows:

[0029]

[0030] st ,

[0031] in, , It is a neural regularity term.

[0032] Furthermore, ,in, intermediate quantity intermediate quantity .

[0033] An autonomous safe navigation system for unmanned agricultural machinery leaving the warehouse based on end-to-end model learning includes:

[0034] The environmental perception module acquires perception information about the unmanned agricultural machinery vehicle body and the surrounding indoor environment;

[0035] The neural network encoding module extracts features from the perceived data, transforming continuous point cloud data into potential distance features that can be used for motion planning.

[0036] The obstacle control safety constraint module establishes safety constraints for the obstacle control function based on potential distance characteristics.

[0037] The model predictive control module designs a learnable model predictive motion controller based on the objective function, potential distance features, and safety constraints. It predicts the future trajectory of the unmanned agricultural machinery based on the current state and point cloud features and generates control commands.

[0038] The execution control module receives control commands output by the model prediction control module and converts them into chassis drive commands to control the unmanned agricultural machinery's movements, enabling autonomous departure operations in an indoor environment.

[0039] The beneficial effects achieved by this invention are as follows:

[0040] (1) A latent distance feature extraction method based on point cloud geometric learning was designed to achieve interpretable safety perception in complex indoor environments. This method enables agricultural machinery to obtain stable, differentiable, and structured geometric distance distribution information, and has significant advantages in terms of computational efficiency, robustness, and geometric topological representation ability.

[0041] (2) A safety constraint based on the control obstacle function with potential distance characteristics was designed to achieve strict safety control across the entire prediction domain. Real-time safety constraints were imposed on the predicted trajectory of the unmanned agricultural machinery to ensure that the machinery maintained the minimum safe distance from obstacles throughout the entire departure process, significantly improving the safety of indoor navigation.

[0042] (3) A learnable model predictive motion controller was designed to achieve integrated collaborative optimization of perception and control. By incorporating neural regularization terms and latent distance features, the motion planning can adaptively adjust the control strategy according to the point cloud geometry. In addition, trajectory accuracy, speed smoothness and safety constraints are jointly considered in the prediction domain, which improves the autonomous departure capability of unmanned agricultural machinery in complex indoor environments.

[0043] (4) An end-to-end integrated software system architecture was designed to achieve consistency and high robustness across the entire link from point cloud perception to control commands. Geometric encoding of point cloud data, inference of potential distance features, construction of control obstacle functions, and optimization of model prediction are all executed in the same computation link, avoiding cross-level information distortion and latency accumulation problems, thereby significantly improving the overall stability of the system. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the autonomous departure safety navigation control system for unmanned agricultural machinery according to the present invention;

[0045] Figure 2 This is a schematic diagram of the unmanned agricultural machinery autonomous departure safety navigation system of the present invention;

[0046] Figure 3 This is a simulation curve of the autonomous departure safety navigation of the unmanned agricultural machinery according to the present invention. Detailed Implementation

[0047] To facilitate understanding and make the objectives, technical solutions, and effects of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the described embodiments are used to explain the methods and systems of this invention and are not intended to limit the scope of protection of this invention.

[0048] This invention provides a method and system for autonomous safe navigation of unmanned agricultural machinery leaving the warehouse based on end-to-end model learning, applicable to various unmanned agricultural equipment. The system achieves high-precision autonomous navigation leaving the warehouse in environments with no satellite positioning signals, confined spaces, and complex structures by integrating a model-based neural network encoder, control obstacle function safety constraints, and a learnable model predictive controller. Its overall control schematic diagram is shown below. Figure 1 As shown.

[0049] The components of an unmanned agricultural machinery autonomous departure safety navigation system based on end-to-end model learning are as follows: Figure 2 The following diagram will be used to specifically illustrate the technical solution of the present invention. The system of the present invention mainly includes the following modules:

[0050] Environmental perception module: This module is responsible for acquiring LiDAR point cloud data of the unmanned agricultural machinery body and the indoor environment in real time, and converting it into a unified vehicle coordinate system. By constructing an environmental representation consistent with the vehicle coordinate system, it can comprehensively reflect the geometric structure of the indoor environment, including the spatial distribution characteristics of walls, columns and dynamic obstacles. This module provides basic data support for subsequent potential distance feature extraction and safety constraint construction.

[0051] The neural network encoding module is primarily responsible for deep feature extraction from the continuous point cloud data acquired by the environment perception module. By employing an optimization problem expansion structure, it transforms the original point cloud into differentiable, structured latent distance features, achieving an end-to-end mapping from high-dimensional dense point clouds to low-dimensional geometric representations. This module provides effective geometric input for subsequent construction of control obstacle functions and model predictive control optimization.

[0052] Obstacle control safety constraint module: This module is used to calculate the geometric safety distance between the unmanned agricultural machine and indoor obstacles in real time based on potential distance features, and to construct safety constraint conditions based on the obstacle control function; by incorporating potential distance features into the safety constraints, this module can constrain the feasibility of the future trajectory of the agricultural machine in the prediction domain, ensuring that the unmanned agricultural machine always operates within the safety set;

[0053] Model Predictive Control Module: This module relies on the kinematic model of agricultural machinery, combined with potential distance features, safety constraints, and optimization objective functions, to construct a learnable model predictive control framework based on the rolling time domain. The neural regularization term introduced in the module can improve the consistency between trajectory planning and geometric structure, enabling the control strategy to adapt to changes in environmental topology. This module ultimately outputs the actual control commands for the front wheel steering angle and speed, and is the core functional module for realizing the autonomous departure behavior decision of unmanned agricultural machinery.

[0054] Execution control module: This module receives control commands output by the model prediction control module and parses and converts them into drive signals that can be executed by the agricultural machinery drive-by-wire chassis, including front wheel steering angle control and drive wheel speed control. This module ensures that the control signals generated by the plan can be executed accurately, so that the unmanned agricultural machinery can accurately complete forward and turning actions according to the control commands and realize autonomous departure from the warehouse in an indoor environment.

[0055] For an embodiment of the above system, an embodiment of an autonomous departure navigation method for unmanned agricultural machinery based on end-to-end model learning includes the following steps:

[0056] S1, based on the autonomous departure requirement of unmanned agricultural machinery from indoor garages to outdoor areas, and with the goal of safe operation of the unmanned agricultural machinery in a lidar-sensing environment, an objective function is established for the indoor autonomous safe navigation of the unmanned agricultural machinery, based on the end-to-end integrated perception and control optimization objective, and state variables and control variables are defined; the objective function is:

[0057]

[0058] Where H represents the total control duration, This represents the attitude information of the unmanned agricultural machinery at various moments, where each moment represents a moment. agricultural machinery posture , Represents coordinates in a local coordinate system. The heading angle representing the unmanned agricultural machinery. This represents the control vector of the unmanned agricultural machinery at each moment. Represent The speed and front wheel angle of the unmanned agricultural machinery are constantly monitored. The feasible region for state and control inputs. and Represent The reference attitude and reference control variables of the unmanned agricultural machinery at all times, Real-time attitude information of unmanned agricultural machinery Represented as:

[0059]

[0060] in, , and The gain matrix can be represented as follows:

[0061] , ,

[0062] in, The wheelbase represents the unmanned agricultural machinery, and the control vector is also mentioned. The safety limits are as follows:

[0063]

[0064] also, and The adaptive weighting coefficients are specifically represented as follows:

[0065]

[0066]

[0067]

[0068] in, and All of these are initial weighting coefficients. , All are adjustment coefficients. The normalized distance between unmanned agricultural machinery and obstacles. For minimum safe distance, for The actual minimum distance between the unmanned agricultural machinery body and the point cloud at any given moment is expressed as follows:

[0069]

[0070] in, for The current LiDAR point cloud collection, Indicates the attitude of agricultural machinery The shape of the vehicle body below, It is the normal vector of each face of the unmanned agricultural machinery. These are the coordinates of a point on the vehicle surface in the vehicle coordinate system. This represents the intercept on the corresponding vehicle body surface. For rotation matrix, and .

[0071] The innovation of this objective function design lies in the fact that the designed objective function is based on an end-to-end integrated perception and control model, which can make real-time adaptive adjustments according to the current perception status information of the unmanned agricultural machinery, so as to achieve efficient obstacle avoidance and tracking control, which is closer to actual engineering.

[0072] S2 uses lidar point cloud information as input and employs a model unfolding method to construct a deep neural network encoder. The original point cloud is generated into a point flow sequence through time-series modeling, and the deep neural network encoder extracts potential distance features to represent the position information of the unmanned agricultural machinery and obstacles. This achieves end-to-end mapping from point cloud data to geometric distance distribution, providing interpretable geometric distance feature input for subsequent control planning.

[0073] The latent distance feature is represented as:

[0074]

[0075] in, , All are optimizations and iterations. The subsequent dual variables represent which part of the agricultural machinery body shape edge participated in the distance calculation with the obstacle point, and the unit vector in the direction of the shortest distance between the obstacle point and the agricultural machinery; The system jointly encodes the distance information from the agricultural machinery to the obstacle point, as well as the geometric matching information of the vehicle's shape. Its calculation form is as follows:

[0076]

[0077] in, This represents the learning rate (step size). For projection operators, Let Lagrange's dual function be the distance from the point cloud to the unmanned agricultural machinery body. This represents the gradient symbol with λ as the independent variable. This represents the gradient symbol with μ as the independent variable.

[0078] The innovation of this neural network encoder module lies in its ability to extract differentiable and structured latent distance features from the original dense point cloud through temporal modeling and optimized unfolding. This approach preserves the true distance topology while maintaining learnability and generalizability. Furthermore, the distance estimation from the point cloud to the agricultural machinery shape is transformed into a trainable optimization process. By using multi-layer unfolding to achieve end-to-end approximate distance field mapping, it effectively avoids the problems of high computational complexity and sensitivity to noise associated with direct minimum distance search.

[0079] S3, for walls, obstacles and dynamic interference targets in complex indoor environments, based on potential distance characteristics, establish safety constraints for the control obstacle function to ensure that the unmanned agricultural machinery maintains a safe distance from obstacles and boundaries during the departure process; the safety constraints for the control obstacle function are:

[0080]

[0081] in, , As slack variables, The control barrier function has the following form:

[0082]

[0083] in, It is a moment An approximate estimate of the minimum distance from the vehicle body to the point cloud, obtained by unmanned agricultural machinery using latent distance features, is calculated in the following form:

[0084]

[0085] st

[0086] The innovation of this control obstacle function module design lies in defining a differentiable safety boundary based on distance characteristics and forcing the system state to remain within a safe set through constraints, ensuring that the future trajectory of the unmanned agricultural machinery satisfies the geometric non-collision condition throughout the entire prediction range. Furthermore, the introduction of the control obstacle function directly transmits the geometric structure of the perception layer to the control optimization layer, enabling the model predictive controller to dynamically adjust the constraint strength based on the point cloud geometric relationships, thereby improving the system's safety margin.

[0087] S4. A learnable model predictive motion controller is designed. Utilizing the kinematic model of the unmanned agricultural machinery, potential distance features, and safety constraints for joint modeling, a neural regularization-based model prediction framework is established. Based on the current state and potential distance features, the future trajectory of the unmanned agricultural machinery is predicted, and control commands are generated, thereby enabling autonomous departure and safe navigation of the unmanned agricultural machinery in complex indoor environments. The learnable model predictive motion controller is as follows:

[0088]

[0089] st ,

[0090] in, , It is a neural regularization term, in the following form:

[0091]

[0092] in, intermediate quantity intermediate quantity .

[0093] The model predictive controller uses a rolling optimization framework to solve for the front wheel steering angle and speed in real time. At each moment... Based on the current state With latent distance features In the prediction time domain The internal solver outputs the optimal control sequence. Take the first item Execute the algorithm, and then recalculate the solution based on the new state in the next moment to achieve continuous tracking and obstacle avoidance in dynamic environments.

[0094] The innovation of this model predictive controller module design lies in its comprehensive consideration of multiple factors within the rolling optimization framework, including trajectory tracking accuracy, speed consistency, control smoothness, geometric consistency, and obstacle safety constraints. Furthermore, the introduction of a neural regularization term directly integrates latent distance features into the optimization objective, enabling the controller to adaptively adjust its control strategy based on the point cloud geometry.

[0095] An embodiment of the autonomous navigation method and system for unmanned agricultural machinery leaving the warehouse based on end-to-end model learning has been specifically explained and illustrated, making the technical solution clearer, more explicit, and easier to implement. Based on this example, simulation experiments were conducted to verify its obstacle avoidance and tracking performance in the absence of satellite signals. In the experiment, a deep unfolded neural network (DNN) was used to process the LiDAR point cloud (generated by the simulator) in real time. The DNN contained six fully connected layers and was trained offline using 100,000 sets of randomly generated point-distance features. The batch size was 256, the learning rate was 5e-5, and a total of 5000 training rounds were performed. After deployment, the unmanned agricultural machinery's body shape was designed according to a medium-scale robot simulator, with a desired speed of 1 m / s. Simultaneously, thousands of ideal point cloud data points within a closed warehouse were generated using the simulator, with the sampling frequency consistent with the actual 2D LiDAR scanning frequency. The path from the starting point to the ending point was approximately 25 meters. In addition, nine randomly generated obstacles, including convex, non-convex, and dynamic ones, were also included. Experimental results show that, using only LiDAR point cloud data, the system can effectively avoid obstacles and achieve tracking control.

[0096] When setting up an unmanned agricultural machine for autonomous navigation out of an indoor warehouse, the indoor obstacle avoidance performance of this embodiment was tested based on the obstacle avoidance results. The experimental results are as follows: Figure 3 As shown, it is easy to see that the end-to-end integrated sensing and control method in this embodiment, under the sensing of lidar, can effectively avoid obstacles and complete tracking control to leave the garage. Therefore, it can be demonstrated that the present invention has excellent indoor autonomous parking navigation capabilities.

[0097] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0098] Although the present invention has been described according to various specific embodiments, those skilled in the art will recognize that the invention can be practiced with modifications within the spirit and scope of the claims. Therefore, any obvious improvements, substitutions, or modifications that can be made by those skilled in the art without departing from the essence of the invention are within the scope of protection of the present invention.

Claims

1. An unmanned agricultural vehicle autonomous off-warehouse safe navigation method based on end-to-end model learning, characterized in that: According to the end-to-end perception and control integrated optimization goal, the target function of the indoor autonomous safe navigation of the unmanned agricultural vehicle is established; The point cloud data acquired by the laser radar, a neural network encoder based on model expansion is established to extract the latent distance features of the point cloud data ; With the potential distance feature as input, the safety constraint of the control barrier function is constructed; Combining the target function, the potential distance feature and the safety constraint, a learnable model predictive motion controller is designed, the future trajectory of the unmanned agricultural vehicle is predicted according to the current state and point cloud feature of the unmanned agricultural vehicle, and control instructions are generated, and the unmanned agricultural vehicle executes the above control instructions.

2. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 1, characterized in that, The objective function is: where H represents the total control duration; represents the pose information of the unmanned agricultural machine at each time, the agricultural machine pose at time , represents the coordinates in the local coordinate system, represents the heading angle of the unmanned agricultural machine; represents the control vector of the unmanned agricultural machine at each time, respectively represent the speed and front wheel steering angle of the unmanned agricultural machine at time is the feasible region of the state and control input, and is the adaptive weight coefficient, represents the reference pose of the unmanned agricultural machine at time represents the reference control amount of the unmanned agricultural machine at time 3. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 2, characterized in that, the unmanned agricultural machine posture information at the time point wherein, , and is a gain matrix, and , , wherein, represents the wheelbase of the unmanned agricultural machine, is a control vector, is a reference control amount that represents the speed of the unmanned agricultural machine at the time point, is a reference control amount that represents the front wheel steering angle of the unmanned agricultural machine at the time point.

4. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 2, characterized in that, , wherein, and are initial weight coefficients, , are adjustment coefficients, is a normalized distance between the unmanned agricultural machine and the obstacle, is a minimum safety distance, is the actual minimum distance between the unmanned agricultural machine and the point cloud at the moment.

5. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 4, characterized in that, wherein, is a set of laser radar point clouds at a moment, represents a vehicle body shape under an agricultural machine posture is a normal vector of each face of an unmanned agricultural machine, is a coordinate of a point on a vehicle body surface in a vehicle body coordinate system, represents a vehicle body surface intercept, and a rotation matrix while .​ 6. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 5, characterized in that, The potential distance feature is: wherein, , are the optimization unfolding iterations the dual variables after the wherein, denotes the learning rate, is the projection operator, is the augmented Lagrangian dual function of the point cloud to the unmanned agricultural machine shape distance, denotes the gradient symbol with λ as the independent variable, denotes the gradient symbol with μ as the independent variable.

7. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 6, characterized in that, The control barrier function safety constraint is: wherein, , is a slack variable, is a control barrier function, which is in the form of: wherein, is the time The unmanned agricultural machine uses the potential distance feature to solve the approximate estimation of the minimum distance from the vehicle body to the point cloud, and the calculation formula is: s.t. .

8. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 7, characterized in that, A learnable model predictive motion controller is: , s.t. , wherein, , is a neural regularizer.

9. The autonomous off-grid navigation method for unmanned farm vehicles from the depot according to claim 8, characterized in that, wherein , intermediate quantity , intermediate quantity .

10. A system for implementing the autonomous off-depot safe navigation method of the unmanned agricultural machine according to any one of claims 1-9, characterized in that, It includes: An environmental perception module acquires the perception information of the vehicle body and the surrounding environment of the unmanned agricultural vehicle; A neural network encoding module extracts features from the perception data and converts continuous point cloud data into potential distance features that can be used for motion planning; A control barrier safety constraint module establishes a safety constraint for the control barrier function based on the potential distance feature; A model predictive control module designs a learnable model predictive motion controller based on the target function, the potential distance feature and the safety constraint, predicts the future trajectory of the unmanned agricultural vehicle according to the current state and point cloud feature of the unmanned agricultural vehicle, and generates control instructions; An execution control module receives the control command output by the model predictive control module and converts it into chassis driving instructions to control the action of the unmanned agricultural vehicle and complete the autonomous off-warehouse operation in the indoor environment.

Images