Livestock pushing robot path planning method based on AI processing

Abstract

The application relates to the field of livestock automation technology and discloses a livestock pushing robot path planning method based on AI processing. Environmental data is collected by using a multi-source perception module, attribute features such as physical density and accumulated volume of forage grass are extracted through a semantic analysis model, and the attribute features are mapped as motion constraint conditions of the robot. A real-time coupling control instruction containing a pose and a push plate deflection angle is generated by an on-board computing unit based on physical field constraints, so that the robot travels along a fluid dynamics path and adjusts the push plate posture in linkage. In addition, the method further comprises an anti-slip correction step based on an optical flow, a crab-shaped approach instruction is generated to offset ground skidding, and a global planning step based on a V2X space-time node, an S-shaped buffer or a tunnel acceleration path is generated to realize continuous operation across regions. The application solves technical problems such as pushing blockage, motor overload, skidding on icy and snowy roads and discontinuous operation across regions in a complex pasture environment.

Description

Technical Field

[0001] This invention relates to the field of livestock automation technology, specifically to a path planning method for livestock feeding robots based on AI processing. Background Technology

[0002] As large-scale farms transform towards intelligent and unmanned operations, automated feed-pushing robots have become key equipment for ensuring dairy cows eat regularly around the clock and increasing milk production. In actual operation scenarios, feed-pushing robots need to perform long-term reciprocating movements in unstructured cowshed passages, based on the geometry of the feed trough edges, to push the scattered feed after the dairy cows have eaten back to the feeding area.

[0003] Existing path planning technologies for pusher robots mainly rely on magnetic navigation, QR code navigation, or laser / vision-based geometric SLAM (simultaneous localization and mapping) technologies. The core logic of these technologies lies in "geometric tracking," that is, using sensors to perceive geometric boundaries in the environment (such as cattle pen lines or walls) and planning a geometric line segment that maintains a fixed distance from the boundary as the travel path. However, in the complex and ever-changing actual working conditions of pastures, this path planning method based solely on geometric constraints has exposed a series of specific technical defects:

[0004] First, existing path planning methods lack the ability to perceive and respond to the physical properties of the work object. The state of hay accumulation in pastures is highly uneven, and different types of hay (such as high-density silage and loose hay) have huge differences in physical resistance characteristics. Existing technologies usually treat all hay as homogeneous geometric obstacles or simple boundary lines. The robot cannot perceive the density and volume potential energy of the hay pile in front. When the robot encounters a sudden high-density hay pile while traveling along a fixed path, it often cannot adjust its path or posture in time, causing the motor output torque to instantly exceed the saturation threshold, resulting in overload shutdown or even mechanical damage. Conversely, in areas with sparse hay, a fixed edge distance may cause the pusher to fail to reach fine feed, resulting in work residue.

[0005] Secondly, existing technologies have significant shortcomings in robustness for kinematic control on low-traction surfaces. Pasture environments are unique, often with snow and ice in winter and wet, slippery manure in summer. Most existing path planning algorithms are based on ideal ground friction models, assuming that the wheel speed integral equals the actual displacement. However, on icy and snowy surfaces, tires are prone to slipping and spinning, causing odometer data to fail. At this time, although the robot receives the command to move in a straight line, the vehicle body will experience uncontrollable lateral slippage under the influence of lateral external forces (such as the reaction force of pushing grass) or ground tilt. Existing geometric tracking algorithms cannot detect this uncontrolled dynamic slippage in real time, let alone generate a reverse compensation trajectory at the planning level, causing the robot to deviate from the predetermined route or even cause a collision.

[0006] Finally, when it comes to global operations involving multiple barns and regions, existing path planning lacks the ability to coordinate across time and space. The physical barriers (such as roller shutters) widely installed in ranches cut the continuous work space into discrete segments. Existing robots usually adopt a passive interaction mode of "perception-deceleration-stopping-waiting-starting". This discontinuous operation mode not only significantly reduces operational efficiency and increases start-stop energy consumption, but more fatally, on icy and snowy roads, frequent starting actions are prone to wheel slippage due to static friction failure, causing the robot to fall into the predicament of "difficulty in starting" in front of the gate.

[0007] In summary, there is an urgent need for a new path planning method that can integrate environmental semantic physical attributes, possess ground dynamics compensation capabilities, and global spatiotemporal node coordination capabilities. Summary of the Invention

[0008] The purpose of this invention is to provide a path planning method for livestock feeding robots based on AI processing, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides a path planning method for a livestock feeding robot based on AI processing, comprising a multi-source sensing module, an onboard computing unit, and a feeding robot with an adaptively adjustable push plate; the method includes: using the multi-source sensing module to collect real-time image data and depth point cloud data of the work area; The method further includes the following steps: Step 1: Process the real-time image data and depth point cloud data using a semantic analysis model to extract the physical attribute features of the forage, which include at least the physical density and the volume of the forage pile. Step 2: Construct a path planning model that includes physical field constraints. The path planning model maps the physical density of forage, the volume of forage pile, and the ground adhesion parameters into motion constraints for the robot. Step 3: Based on the path planning model, the onboard computing unit calculates and generates coupled control commands that include the robot's pose and the pusher deflection angle in real time. Step 4: In response to the coupling control command, the pushing robot moves along the generated hydrodynamic path and dynamically adjusts the deflection angle of the pusher plate, so that the deflection angle of the pusher plate is adjusted in conjunction with the change of the physical property characteristics.

[0010] Furthermore, the path planning model employs a nonlinear model predictive control architecture. The execution steps of the nonlinear model predictive control architecture include: Establish a system prediction model that incorporates the dynamic characteristics of the robot chassis and the nonlinear interaction mechanism between forage and grass; Define the system state vector and control input vector, wherein the control input vector includes at least the drive acceleration, the front wheel steering angle, and the push plate angle adjustment angular velocity; Within the prediction time domain, the future state of the robot is iteratively deduced based on the system prediction model. Under the premise of satisfying the physical characteristics constraints of the motor and the ground adhesion constraints, the optimal control input sequence that minimizes the preset multi-objective optimization function is solved.

[0011] Furthermore, the mathematical expression of the multi-objective optimization function includes a trajectory tracking error term, an energy consumption cost term, and a control increment penalty term; The trajectory tracking error term is used to characterize the Euclidean distance between the robot's current position and the ideal work line; The energy consumption cost item is calculated based on the motor output driving force, which is derived through the dynamic balance equation. The control increment penalty term is used to limit the rate of change of acceleration, steering angle, and push plate angle.

[0012] Furthermore, the system prediction model includes logic for calculating forage pushing resistance, and the calculation of pushing resistance follows the following:

[0013] in, For the resistance of pushing material, This is the forage deformation resistance coefficient. Forage physical density, Forage pile volume, This represents the current deflection angle of the push plate. The coefficient of kinetic friction between the forage and the ground. The positive pressure generated by the weight of the forage.

[0014] Furthermore, the path planning model employs a fuzzy inference system, and the execution steps of the fuzzy inference system include: A fuzzy input model is established to map the physical density of the forage into a fuzzy variable of the forage quantity level, and to map the current feedback value of the pushing robot into a fuzzy variable of the load rate. Construct a fuzzy rule base, which contains multiple logical rules to define the path offset and push plate deflection angle under different combinations of grass level and load rate; The fuzzy inference engine is used to perform calculations on the input fuzzy variables to output a fuzzy control set; the centroid method is used to defuzzify the fuzzy control set to obtain accurate path offset values ​​and push plate angle values.

[0015] Furthermore, the final generation of the fluid dynamics path follows a correction logic based on the rate of change of deviation, and its calculation follows the following:

[0016] in, The target path coordinates, Using the coordinates of the baseline line, For fuzzy output gain coefficient, The path offset values ​​obtained for defuzzification. This is the integral compensation coefficient. For real-time load rate, The target load rate.

[0017] Furthermore, the method also includes a correction step based on optical flow anti-slip: Ground texture features are collected using a downward-looking visual sensor, and the average pixel movement speed of the ground is calculated using the dense optical flow method. The average pixel movement speed is converted into physical observation speed by combining the camera intrinsic parameter matrix; The difference between the physically observed velocity and the theoretical velocity fed back by the robot wheel speedometer is calculated to obtain the actual lateral slip velocity vector; An anti-slip compensation force vector, which is opposite in direction to the actual lateral slip velocity vector, is superimposed on the path planning model.

[0018] Furthermore, the step of generating coupled control commands including robot pose and pusher deflection angle also includes generating crab approach control commands: Based on the anti-slip compensation force vector and the resultant force vector of the path planning model, the crab-shaped compensation angle is calculated. The control of the robot's chassis yaw angle deflects the crab-shaped compensation angle, so that the robot's front direction forms an angle with the actual speed direction, and the longitudinal component of the wheel rolling friction force is used to counteract lateral slippage.

[0019] Furthermore, the method also includes a global planning step based on V2X spatiotemporal nodes: The on / off status and action response time of the pasture's physical barriers are obtained through the V2X communication link; Establish a spatiotemporal node potential energy function for the physical partition facility in the global map; The spatiotemporal node potential energy function includes a static repulsive term based on spatial distance and a time modulation term based on the facility's activation progress. The value of the time modulation term decreases as the facility opening height increases.

[0020] Furthermore, the global planning step based on V2X spatiotemporal nodes also includes action chain generation logic: Based on the time difference of arrival reverse reasoning strategy, the ideal arrival speed of the robot to the physical partition facility is calculated; When the facility is not fully open and the robot will arrive earlier than the opening time if it travels in a straight line, a high-damping S-shaped buffer path is generated to control the robot to travel along the S-shaped trajectory to consume time. When the facility's opening level reaches the passage threshold, a low-damping tunnel straight path is generated, and the environmental damping coefficient is adjusted to control the robot to accelerate through.

[0021] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention breaks through the limitations of traditional geometric tracking by extracting the density and volume characteristics of forage through AI semantic analysis and mapping them into a multi-dimensional potential energy field or dynamic constraint. The robot can automatically plan a fluid-like dynamic path based on the magnitude of the load's "potential energy" (e.g., automatically generating an S-shaped cutting trajectory when encountering a large haystack), and couple the pusher deflection angle in real time. This "load-based path determination" mechanism avoids motor overload and mechanical damage caused by forcefully pushing high-density haystacks, while ensuring thorough cleaning in sparse areas, significantly improving the process quality and equipment lifespan.

[0022] 2. This invention introduces an anti-slip compensation mechanism based on optical flow perception and a crab-like approach control strategy. By monitoring the texture movement vector of the ground in real time, it accurately eliminates false displacements caused by tire slippage and actively generates a reverse compensation force field and sideslip angle commands at the path planning level. This proactive defense strategy of "walking in a straight line at an angle" effectively counteracts lateral slip interference caused by low-traction surfaces, ensuring that the robot can maintain centimeter-level path tracking accuracy even in extremely cold, snowy, or slippery environments, achieving all-weather, fault-free operation.

[0023] 3. This invention transforms physical barriers into dynamic spatiotemporal nodes within the overall geodynamic energy field. Through action chain generation mechanisms (such as S-shaped buffer paths or tunnel acceleration modes), it achieves non-contact collaboration between robots and infrastructure. Robots can dynamically adjust their movement rhythm according to the opening progress of doors, traversing barriers in a "flowing" manner. This avoids high-risk, high-energy-consumption start-stop actions, significantly improving the continuity and overall energy efficiency of unmanned operations across the entire ranch. Attached Figure Description

[0024] Figure 1 This is a flowchart of the fluid path planning method based on a multidimensional potential energy field according to the present invention. Figure 2 This is a flowchart of the dynamic programming process based on nonlinear model predictive control according to the present invention. Figure 3 This is a logic block diagram of the biomimetic path planning based on the fuzzy reasoning system of the present invention; Figure 4 This is a schematic diagram of the control logic for introducing optical flow anti-slip correction in this invention; Figure 5 This is a flowchart of the global path planning process based on V2X spatiotemporal node collaboration in this invention. Detailed Implementation

[0025] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1-5 This invention provides a path planning method for livestock feeding robots based on AI processing; Example 1 like Figure 1 As shown in the figure, this embodiment discloses a fluid path planning method for a pasture pushing robot based on a multi-dimensional potential energy field. This method aims to solve the core technical problems such as obstruction of the pushing robot's operation, motor overload, and incomplete pushing caused by uneven forage accumulation and changes in ground adhesion in complex unstructured pasture environments.

[0027] The core idea of ​​this solution is to construct the robot's motion space as a virtual force field constrained by multiple physical fields. The robot is regarded as a point mass moving in this force field, and its trajectory is no longer a pre-set geometric line segment, but a fluid dynamic path generated based on the real-time perceived environmental potential energy distribution. The control system first collects real-time image data and depth point cloud data of the working area through high-resolution intelligent cameras and depth cameras mounted at the front end. It then uses the built-in intelligent brain to perform deep fusion processing on the multi-source data. In order to achieve accurate quantification of unstructured forage morphology, this embodiment establishes a semantic segmentation and physical attribute mapping method based on an improved fully convolutional neural network. The model is based on a massive dataset collected from pastures, encompassing image samples of various forages such as silage, hay, and alfalfa under different lighting and stacking conditions. During model training, the input layer receives normalized RGB images and depth maps. An encoder extracts multi-scale features, and a decoder restores spatial resolution. The output layer generates a forage distribution probability map through pixel-level classification. To accurately characterize the physical resistance properties of forage, the system introduces a texture entropy analysis algorithm. By calculating the energy and entropy values ​​of the image's gray-level co-occurrence matrix, visual texture features are transformed into a physical density index of forage. Specifically, the system defines the physical density of forage. It exhibits a non-linear mapping relationship with image texture features, and the estimated density value of the current region is obtained through weighted calculation. This density value directly reflects the potential mechanical resistance during the material pushing operation.

[0028] Simultaneously, by combining the spatial geometric information acquired by the depth camera, the system calculates the accumulation volume of forage in real time. This perception mode, which combines deep learning with physical feature extraction, provides precise quantitative input for the subsequent construction of the potential energy field, ensuring that the robot can not only see obstacles but also understand their physical properties. After obtaining accurate environmental semantic information, the core step of this embodiment is to construct a composite artificial potential energy field that includes a gravitational field, a repulsive field, and a damping field. Unlike the traditional path planning method that only considers geometric distance in potential energy field construction, this scheme proposes a dynamic potential energy field model based on load prediction.

[0029] First, construct a gravitational field pointing towards the target line of the hopper edge. The gravitational field generates a continuous attractive force, keeping the traction robot close to the edge of the trough to prevent missed pushes. The gravitational potential energy function is mathematically defined as a quadratic function of the Euclidean distance between the robot's current position and the target work line, generating a gravitational force. The magnitude is proportional to the distance the robot deviates from the target line, that is:

[0030] in This is the gravitational gain coefficient. For distance deviation, the physical meaning of this formula is that when the robot is too far from the feed trough, the gravitational force increases linearly, forcing the robot to quickly return to the working baseline; while when the robot is close to the feed trough, the gravitational force gradually decreases, allowing the robot to make fine adjustments within a local range. However, a gravitational field alone will cause the robot to ignore obstacles in front of it and move in a straight line. Therefore, a repulsive field based on the distribution of forage must be constructed. This is the key innovation of this embodiment. Traditional repulsive fields are only related to the distance to obstacles, while the repulsive field strength in this solution is related to the forage density determined by AI semantic segmentation. and volume They show a positive correlation.

[0031] To quantify this physical process, the system defines the repulsive field gain coefficient. As a function of density and volume, its calculation logic follows the formula:

[0032] in , , All are preset weighting factors. The exponential term in this formula indicates that as the density of the forage increases, the repulsion gain coefficient increases exponentially. The physical meaning of this mathematical design is very clear: when facing loose hay, density... With a smaller density and lower repulsive gain, the robot experiences a weaker "repulsive force," allowing path planning to enable it to approach the haystack for work. However, when faced with high-density compacted silage, the density... As the repulsive force increases, the repulsive gain rises sharply, generating a powerful virtual repulsive force. The repulsive force The calculation formula further incorporates a distance factor, namely:

[0033] in The real-time distance between the robot and the heavily loaded haystack. The critical influence distance for the repulsive force is defined as follows: when the robot enters this critical influence distance and detects a high-density haystack, the strong repulsive force is superimposed on the attractive force, causing the resultant force to point away from the cattle pen. This allows the robot to plan a smooth, outward-convex bypass trajectory. This mechanism perfectly simulates the flow phenomenon when a fluid encounters a highly damped object, fundamentally avoiding motor overload or mechanical damage caused by the robot forcibly cutting into the high-density haystack. In addition to addressing the spatial constraints caused by the hay load, this embodiment also fully considers the influence of the pasture ground environment on the robot's motion control. Especially on icy or wet roads in winter, the drastic changes in the ground friction coefficient can severely affect the stability of path tracking. Therefore, this solution introduces a ground damping field into the potential energy field model. .

[0034] The system collects real-time data from the wheel speed sensors and the actual acceleration data from the inertial measurement unit (IMU) of the automotive-grade drive-by-wire chassis, and estimates the current ground slip rate using an extended Kalman filter algorithm. When the ground slip rate Exceeding the preset security threshold When the system determines that it is currently on a low-adhesion surface, it immediately activates the high-damping mode. The damping field works by limiting the rate of change of curvature of the path, that is, generating a reverse damping torque on the robot's angular velocity. The damping force is defined by the system. With slip ratio and the current path curvature It exhibits a non-linear proportional relationship, and its mathematical model can be expressed as:

[0035] in The damping adjustment coefficient, the physical meaning of this formula is that the more slippery the ground ( The larger the curve (or the sharper the bend in the planned path) When the magnitude of the damping force increases, the damping force increases quadratically. This virtual damping force is added as a penalty term to the gradient descent calculation of the total potential energy field, forcing the generated planned path to become gentler and avoiding sharp turns on icy and snowy roads. This prevents the vehicle from losing control or skidding due to excessive lateral force. Through the organic superposition of the gravitational field, repulsive field and damping field, the environment in which the robot is located is digitized into a continuously changing potential energy surface.

[0036] The robot's planned path is the streamline along the direction of the fastest decrease in potential energy on this surface, with the resultant force vector as follows:

[0037] This embodiment provides real-time guidance for the robot's next movement. After generating a highly adaptive fluid path, it further achieves deep coupling control between the actuator's posture and the path's geometry to ensure the quality of the material pushing operation. The core actuator of the pushing robot is the adaptively adjustable push plate, whose deflection angle... The control is no longer an independent open-loop logic, but rather related to the tangential direction of the planned path. The system establishes a real-time function mapping relationship between the repulsive field strength and the target deflection angle of the push plate. Following the coupling control law, the mathematical expression of this control law is:

[0038] In the formula, From the perspective of basic assignments, and These are the force field coupling coefficient and the curvature coupling coefficient, respectively. The physical meaning of this mathematical model is extremely profound: the second term in the formula... This characterizes the ratio of the current encountered hay load to the repulsive force when the robot encounters a high-density hay bale. When the number of surges increases dramatically, this value increases significantly, driving the push plate angle. The robot automatically deflects significantly outward, creating a larger angle between the pusher and the direction of travel. This transforms the original forward pushing action into a lateral cutting action, using the decomposed force of the robot's movement to gently "shave" the haystack back into the chute, forming a composite motion trajectory similar to an "S-shaped cut." (The third term in the formula...) This represents the rate of change of the angular velocity of the path, i.e. the curvature of the path. When the robot performs a circumvention maneuver, the change in the curvature of the path generates a feedforward signal, which adjusts the angle of the push plate in advance to match the body posture, ensuring that the push plate is always close to the curve of the feed trough or the edge of the hay pile, thus avoiding mechanical interference or hay leakage.

[0039] This dual coupling mechanism of "path-attitude" enables the robot to overcome resistance like a skilled driver when facing sudden high-load conditions, through a combination of "steering and side-shoveling" actions. This achieves a leap from rigid geometric tracking to intelligent physical adaptation. Through the above complete technical solution, this embodiment not only solves the adaptability problem of traditional pusher robots in complex pasture environments, but also integrates environmental perception, path planning and execution control through a rigorous mathematical model, ensuring the robustness and reproducibility of the technical solution in practical engineering applications. Based on the above-disclosed algorithm model and control logic, those skilled in the art can fully construct a pasture pusher robot system with the same level of intelligence.

[0040] Example 2 like Figure 2As shown, this embodiment discloses a high-performance dynamic path planning method for a pasture feeding robot based on nonlinear model predictive control. This method aims to solve the technical challenges of traditional potential field algorithms easily getting trapped in local minima and failing to accurately handle multiple physical constraints in complex pasture operation scenarios with ample computing resources and extremely high dynamic constraints. The core of this solution lies in no longer relying solely on current sensing data to generate instantaneous reactive control commands, but instead introducing a nonlinear model predictive control (NMPC) architecture with predictive capabilities. By establishing a system state equation that includes the robot chassis dynamics and the nonlinear interaction mechanism between forage and forage, the optimal control sequence that satisfies all physical constraints and minimizes the comprehensive cost function is found within a finite prediction time domain. This control system first obtains the robot's pose state and environmental semantic information at the current moment through the front-end sensing module, and then enters the core solution stage of the model predictive controller. To accurately describe the complex physical interaction between the robot and forage, this embodiment constructs a system prediction model integrating kinematics and dynamics. In this model, the robot's state vector is defined as:

[0041] in and Let be the robot's position coordinates in the ranch's global coordinate system. For heading angle, For the longitudinal driving speed, the control input vector is defined as:

[0042] in To drive acceleration, This refers to the front wheel steering angle. The angular velocity for adjusting the push plate angle follows a discrete-time kinematic equation, i.e., the state at the next moment. The current state With control input The function is denoted by kinematic equations alone; however, kinematic equations alone are insufficient to describe the load changes during the feeding process. Therefore, this scheme deeply integrates the forage resistance dynamics equation into the model, and the feeding resistance is defined by the system. It exhibits a highly nonlinear correlation with the physical properties of forage and the pushing posture, and its calculation formula is as follows:

[0043] In the formula, This is the forage deformation resistance coefficient. The physical density of forage as identified by the visual semantic model. This refers to the volume of forage piled up in front of the pusher. This represents the current deflection angle of the push plate. The term indicates that the magnitude of the resistance is directly proportional to the sine of the angle of attack of the push plate. The coefficient of kinetic friction between the forage and the ground. The physical meaning of this dynamic model, which considers the positive pressure generated by the gravity of the forage, is that when the robot faces high-density silage or large-volume haystacks, the pushing resistance will increase significantly, and this resistance can be adjusted by changing the angle of the pusher plate. To perform nonlinear adjustment.

[0044] After constructing an accurate system prediction model, the core of this embodiment lies in solving for the optimal control sequence using a rolling time-domain optimization algorithm. This process relies on a carefully designed multi-objective optimization function. The objective function aims to balance the trade-offs between material feeding efficiency, operational energy consumption, mechanical wear, and path smoothness. Its mathematical expression is as follows:

[0045] In the formula, The step size for prediction in the time domain represents the length of the future time window that the system predicts forward. The first term... This is the trajectory tracking error term, used to quantify the Euclidean distance between the robot's current position and the ideal feed trough edge work line. The weight of this term... The amount is relatively large to ensure the process requirement of "minimum residue after material pushing," the second item. This is the energy consumption cost item, of which The driving force output by the motor is determined by the dynamic equilibrium equations:

[0046] Calculations show that this measure aims to suppress excessive torque output from the motor and prevent energy waste caused by "overloaded hard pushing." The third measure... To control incremental penalty terms, which limit drastic changes in acceleration, steering angle, and push plate angle, ensuring extremely smooth generated paths and preventing impact on the mechanical structure, the fourth term... The safety potential energy penalty term, based on obstacle repulsion, increases exponentially when the predicted trajectory approaches an obstacle or cattle pen, forcing the optimizer to abandon the candidate solution for that path. When solving for the minimum of the above objective function, the system must strictly adhere to a series of hard physical constraints to ensure the robot's operational safety. The first constraint is the physical characteristics of the motor, specifically the motor output torque. It must be less than or equal to the motor's maximum peak torque. ,Right now .

[0047] When encountering an exceptionally large haystack, leading to predictions... When the required torque exceeds this threshold, the optimizer will automatically adjust the control input, for example, by reducing the speed. Or increase the deflection angle of the push plate To reduce the load and find a feasible solution within the constraints, the system first considers ground adhesion constraints. Secondly, to prevent sideslip on wet surfaces, a lateral slip ratio constraint is introduced. , of which lateral slip ratio The calculation relies on real-time comparison between visual optical flow feedback and the vehicle dynamics model. When the predicted steering action may cause the slip ratio to exceed the critical value, the system will forcibly limit the maximum steering angle. To ensure the robot's stability on icy and snowy roads, this embodiment employs an advanced rolling time-domain solution strategy to achieve the above optimization process. Within each control cycle, for example, every 50 milliseconds, the controller reads the current state observation and uses this as a starting point to iteratively deduce within a 2-second prediction window. The optimization algorithm (such as Sequential Quadratic Programming (SQP) or the interior-point method) searches for a set of algorithms that satisfy all dynamic constraints and boundary conditions to achieve the objective function. Control input sequence to reach global minimum This optimal sequence contains the best speed, best turning angle, and best pushing angle for future moments. However, to cope with dynamic environmental disturbances and the inherent uncertainties of the model, the robot does not blindly execute the entire sequence, but only executes the first instruction in the sequence. This refers to the optimal control quantity at the current moment. When the next control cycle arrives, the system will update the starting point of the state based on the latest sensor feedback data and re-perform prediction and optimization for the entire cycle.

[0048] This "take one step, observe one step, calculate multiple steps" feedback correction mechanism gives the material-pushing robot a strong ability to adapt to its environment. For example, when a high-density haystack suddenly appears in front of the robot caused by cattle, the predictive model will immediately calculate the motor torque loss if it continues to move in a straight line. Exceeding the saturation threshold This leads to a surge in energy consumption terms and constraint penalty terms in the objective function, in order to minimize the total cost. The optimizer will quickly converge to a new set of solutions: that is, instructing the chassis to fine-tune the steering outwards (changing...) At the same time, the instruction push plate rapidly increases the deflection angle (changes) This process forms a smooth, circular cutting trajectory. This process is not based on a preset rule base, but on the mathematically optimal solution calculated in real time by the physical model. In addition, this method is naturally able to handle multivariate coupling problems. While adjusting the path, it automatically matches the optimal pusher angle and travel speed, realizing integrated optimal coordination from path planning to actuator control. Through the nonlinear model predictive control scheme provided in this embodiment, the pushing robot can not only operate safely at the edge of dynamic limits, but also achieve the optimization of system energy efficiency while ensuring the quality of pushing. It completely solves the problem of oscillation or getting trapped in local optima that may occur in the potential energy field algorithm in complex and confined spaces. It provides solid algorithmic support for realizing the ultimate unmanned and intelligent pushing operation in pastures. Based on the above-disclosed system state equations, objective function construction principles and constraint settings, those skilled in the art can fully reproduce this control method and apply it to actual livestock automation equipment by combining it with existing embedded high-performance computing platforms.

[0049] Example 3 like Figure 3 As shown, this embodiment discloses a low-computing-cost biomimetic path planning method based on a fuzzy inference system. This method is specifically designed for the application scenarios of push-type robots in animal husbandry, which are widely used in the industry and are characterized by cost sensitivity, limited computing resources, and highly unstructured operating environments. Unlike the schemes based on high-dimensional potential energy fields in Embodiment 1 or model predictive control in Embodiment 2, this embodiment no longer relies on precise analytical geometric modeling of the environment or solving high-order differential equations. Instead, it establishes a set of fuzzy control logic that mimics the driving experience of human operators (i.e., "experienced drivers"). By converting fuzzy perception information into deterministic control quantities, real-time path adaptation and load management on a low-cost embedded controller are achieved.

[0050] The core of this technical solution lies in constructing a closed-loop control system that includes a fuzzy interface, an expert rule base, a fuzzy inference engine, and a defuzzification interface. This solves the problems of sudden action changes, oscillations, and poor adaptability that traditional threshold logic control encounters when facing complex working conditions, enabling smooth operation of the pushing robot under nonlinear load variations. This embodiment first establishes a fuzzy input model that can convert precise sensor physical quantities into qualitative linguistic variables, which is the foundation for realizing "bionic perception." The system defines two core input variables: visually perceived forage quantity levels. and chassis-sensed current load rate For input variables The data originates from the image processing results of the front-end camera. The system does not directly use the raw value of pixel ratio, but instead maps it to the domain of discourse. Above, four subsets of fuzzy language were defined:

[0051] To describe these fuzzy concepts, the system establishes a membership function based on a Gaussian distribution. It describes the degree to which a specific input value belongs to a particular fuzzy set, and its mathematical expression is:

[0052] In the formula, The center point of the membership function represents the most typical value of the linguistic variable; The standard deviation determines the width and sensitivity of the function. The physical meaning of this formula is that sensor readings are no longer binary logic, but rather... There are probability values ​​between them. For example, when the amount of forage is in the transition zone between "moderate" and "a lot", it may belong to both sets at the same time, but with different membership degrees (e.g., 0.6 belongs to moderate and 0.4 belongs to a lot). This allows the control system to smoothly handle boundary conditions and avoid control jumps caused by "threshold" judgment.

[0053] Similarly, for input variables The system collects the real-time current feedback value of the servo driver. And normalize it to the load factor:

[0054] in This is the no-load current. The critical current for motor stall is... On the domain of discourse, fuzzy subsets are defined:

[0055] It also uses Gaussian or trapezoidal functions for fuzzy mapping, which allows the robot to perceive subtle state differences like humans, such as "the load is a bit heavy but it can still be pushed" or "it can't be pushed at all".

[0056] After fuzzifying the input information, this embodiment constructs a fuzzy rule base based on expert experience knowledge. This is the core of the entire system's decision-making process. The rule base uses "IF-THEN" format language control rules to simulate the intuitive reactions of skilled operators facing different working conditions. The system establishes a... The multidimensional rule matrix covers 16 typical operating conditions, ranging from "sparse grass and no load" to "abundant grass and overload." For example, one typical obstacle avoidance rule is defined as follows:

[0057]

[0058] Here Represents the path offset. The deflection angle of the pusher plate represents the physical meaning of this rule: when the robot "sees" a large haystack ahead and its body "feels" that the motor is about to stall, it should immediately plan a path that deviates significantly from the cattle pen (significantly outward), while simultaneously adjusting the pusher plate angle to its maximum (forming a cutting posture), sacrificing the amount of material pushed in a single pass in exchange for improved maneuverability. Another rule is:

[0059]

[0060] The physical meaning is: when the hay is sparse and offers almost no resistance, the robot should move inward towards the cattle pen, keeping the pusher vertical to ensure that the fine hay is thoroughly cleared. For mathematical reasoning, the system uses the Mamdani inference method, determining the activation strength of each rule through a "minus" operation. Assuming the membership degree of the input hay quantity at a certain moment is... The load membership degree is The output membership degree of this rule This is the minimum membership degree of the predecessor, i.e.:

[0061] This logical operation ensures the conservatism and safety of control decisions, meaning the system's response is limited to the most critical sensing condition, avoiding aggressive actions caused by false alarms from a single sensor. The inference engine output is a truncated fuzzy set (graphic), which must be converted into precise numerical commands to control the actuators—a process known as defuzzification. In this embodiment, the centroid method is preferably used to calculate the final control output. This method finds the optimal balance control point by calculating the geometric centroid of the region enclosed by the fuzzy output surface and the coordinate axes. The mathematical formula is as follows:

[0062] In the formula, To provide accurate output values ​​after deblurring (such as specific path offset in meters or push plate angle in degrees). To output the elements in the universe of discourse of the output variable, The membership function of the aggregated fuzzy set is the final output value. The physical meaning of this formula is that the final output value... It is a weighted average of all activated rules, representing a "compromise" and "balance" for the current complex operating conditions. For example, when the rule suggestions "slightly outward" and "significantly outward" are both partially activated simultaneously, the centroid method automatically calculates a smooth value between the two, thus achieving continuity in path planning based on the calculated path offset. The system generates the final target travel path coordinates. To prevent overshoot or oscillation during the control process, the system introduces a correction term based on the rate of change of deviation. The final path generation formula is described as follows:

[0063] in, The reference straight line along the edge of the trough. For fuzzy output gain coefficient, The integral compensation coefficient, in its physical sense, means that the robot's actual travel path is based on a reference straight line, superimposed with an avoidance offset calculated by fuzzy logic, and an integral term based on load current deviation is introduced to eliminate steady-state error. When the system detects a current load... consistently above the target value In this process, the integral term gradually increases the path offset over time, forcing the robot to move further away from the haystack until the load returns to the normal range. This design ensures that the robot can not only cope with instantaneous load changes, but also adapt to long-distance, high-density forage operation environments, reflecting the biomimetic intelligence of "load-based path determination".

[0064] Example 4 like Figure 4 As shown, as a deeper exploration and expansion of the core solution described in Example 1, this example discloses a highly robust path planning method for a pushing robot that introduces an "optical flow anti-slip" potential energy field correction mechanism. It specifically addresses the technical pain point of Example 1, where the potential energy field-planned path cannot be accurately executed due to the sudden drop in the ground friction coefficient on low-adhesion road surfaces such as extremely cold ice and snow, and wet and slippery manure. On conventional roads, the robot's kinematic model assumes that there is sufficient static friction between the tires and the ground to provide lateral grip in response to steering commands. However, in the winter scenario of a pasture, this assumption often fails. Although the robot receives the "detour" command generated by the potential energy field, the robot body is prone to uncontrollable lateral slippage under inertia, causing the actual trajectory to deviate from the planned equipotential line, which in turn leads to safety accidents such as collisions with cattle pens or falling into manure ditches.

[0065] To completely solve the problem of "path tracking failure under nonholonomic constraints," this solution builds a closed-loop defense system on top of the original potential energy field architecture. This system includes a visual optical flow slip perception model, an anti-slip force field compensation model, and a crab-shaped approach kinematic control model. This embodiment first establishes a real-time ground slip perception model based on dense optical flow. This model aims to address the deficiency of traditional wheel speedometers in accurately reflecting the robot's true displacement through integration during slippage and idle. The system mounts a set of high-frame-rate, downward-looking monocular industrial cameras on the robot's underside. These cameras are equipped with a high-intensity LED array to ensure clear capture of ground texture features even at night or in shaded areas. The core of the perception model is based on the assumption of constant brightness, i.e., the brightness of the same pixel in the image... In an extremely short time interval The interior remains unchanged, although its spatial position has shifted. Based on Taylor series expansion and neglecting higher-order terms, the optical flow constraint equations were systematically constructed:

[0066] in, and These represent the spatial gradients of the image grayscale in the horizontal and vertical directions, respectively. For time gradient, and This refers to the instantaneous velocity vector in the pixel coordinate system to be solved. Since the equation for a single pixel cannot solve for two unknowns (aperture problem), this embodiment uses the Farneback dense optical flow algorithm. By establishing a polynomial expansion model in the local neighborhood of the image, the least squares method is used to perform weighted regression on the motion vectors of all feature points in the field of view, thereby calculating the average pixel movement velocity of the entire ground. In order to convert the pixel velocity into a physically meaningful actual physical velocity, the system introduces the camera's intrinsic parameter matrix. and installation height Perform perspective transformation projection, assuming the camera focal length is... Then the observation velocity in the ground physical coordinate system With pixel speed The relationship exhibits a linear proportional characteristic, that is:

[0067] At this time, the system synchronously acquires the theoretical linear velocity fed back by the chassis wheel speed encoder. In an ideal no-slip state, It should be strictly equal to However, on icy and snowy roads, the two will deviate significantly. The system defines the actual lateral slip velocity vector. Let be the vector difference between the visually observed speed and the theoretical wheel speed, and its mathematical expression is:

[0068] in The rotation matrix represents the robot's current heading angle. The physical meaning of this formula lies in directly measuring the relative motion to the ground through vision, an "external observer," thereby eliminating the false displacement caused by tire slippage and accurately quantifying the uncontrolled sideslip component driven by inertia, and precisely obtaining the lateral slip velocity. Subsequently, this embodiment does not simply use it as an alarm threshold, but instead introduces it as a negative feedback signal into the artificial potential energy field model established in Embodiment 1. By constructing an "anti-slip compensation vector," active intervention at the path planning level is achieved. In the original scheme A, the robot's planned path is determined by the resultant force vector of the gravitational and repulsive fields. In this embodiment, a dynamic anti-slip compensation force vector is superimposed on the system. Its direction is always related to the detected slip velocity. Instead, the system aims to generate a virtual "corrective force." The magnitude of this compensating force is not a simple linear relationship. To cope with severe instability on roads with extremely low friction coefficients, the system employs a nonlinear gain adjustment strategy, and the compensating force vector... The mathematical model is defined as follows:

[0069] In the formula, Based on the compensation gain coefficient, It is a nonlinear radical factor. The term indicates that the compensation force is directly proportional to the square of the slip velocity. The physical meaning of this formula is extremely crucial: when the slip velocity is small, the system applies a linear corrective force for minor adjustments; while when the slip velocity is large... When the force increases dramatically (e.g., a vehicle suddenly enters an icy area), the compensating force will grow exponentially, rapidly generating a strong counter-component in the potential energy field. This counter-component will significantly alter the resultant force vector. The direction of the slide forces the path planner to generate an "over-pre-aimed" trajectory in the opposite direction of the slide. This correction mechanism essentially introduces dynamic feedforward at the path planning level. It tells the robot: "Since you will slide to the left, I will plan a path that deviates to the right. Your slide error will just offset the path deviation, so that the final synthetic trajectory will still move in a straight line along the original edge of the trough." However, making corrections at the path level alone may not be enough for a pusher robot with huge inertia, because the wheels may have lost most of their lateral grip on the ice.

[0070] To address this, this embodiment further proposes a special kinematic control strategy based on "crab approach," which directly controls the chassis attitude angle through active yaw control. Traditional differential or Ackerman steering models assume that the direction the vehicle's head is pointing is the velocity direction, but in this scheme, the system allows for an angle between the vehicle's heading and the actual velocity direction, i.e., the crab angle. The system is based on the corrected resultant force vector. Decomposed expected longitudinal component and expected lateral component Real-time calculation of the required crab-shaped compensation angle Its calculation formula follows the arctangent relationship:

[0071] in This is the sideslip damping coefficient, used to suppress vehicle sway. The physical meaning of this formula is that in order to generate sufficient lateral resistance to counteract the sliding force on the ice, the robot must actively steer its front end at an angle in either the "uphill" or "anti-slip" direction. The control system utilizes the longitudinal component of the wheel's rolling friction to provide the lateral support force that would otherwise be provided by static friction, and adjusts the target angle accordingly. The input to the yaw angle closed-loop controller of the chassis drives the left and right wheels to generate differential speed, so that the robot maintains a "tilted forward" posture. Although this posture makes the front of the robot appear tilted, its actual center of mass trajectory can strictly follow the potential energy field equipotential lines planned in scheme A, thus achieving physical cancellation of lateral slippage. In order to ensure the safety and stability of the entire optical flow anti-slip system, this embodiment also introduces a set of rigorous critical safety boundary judgment logic. Although "crab approach" and "potential energy field correction" can cope with a certain degree of slippage, the laws of physics determine that friction has a limit. The system defines a slippage loss control critical value. This threshold is set based on the limiting dynamic friction coefficient between the tire and the ice surface. and vehicle quality According to the physical formula:

[0072] To make an estimate, among which To determine the maximum permissible safe buffer distance, the system monitors the sliding velocity modulus measured by optical flow in real time. When detected When the slip exceeds the physical limits of dynamic compensation, the system no longer attempts to recover through path correction but immediately triggers the "ABS intermittent braking" deceleration logic or emergency braking logic to prevent vehicle spin-out due to forced correction. Furthermore, to prevent noise interference from the optical flow sensor in areas of specular reflection (such as puddles), the system incorporates a Kalman filter. Smoothing is performed, and the noise covariance matrix is ​​observed. The system adaptively adjusts based on image texture richness. When the texture richness is below a preset threshold (such as a pure ice surface or complete darkness), the system automatically reduces the confidence weight of visual optical flow and smoothly transitions to a short-time dead-counting mode based on IMU inertial navigation, ensuring that the control system does not experience violent oscillations at the moment of perception failure.

[0073] Example 5 like Figure 5 As shown, this embodiment discloses a method for global potential energy field management and dynamic path planning based on V2X spatiotemporal node collaboration. As a deep extension and technological upgrade of the local micro-path planning scheme described in Embodiment 1 in the macro-level global operation dimension, this scheme aims to completely solve the technical problems of potential energy field breakage, unnecessary start-stop of robot operation, surge in energy consumption, and slippage when the pushing robot operates across cattle sheds and areas due to the discontinuous opening characteristics of physical isolation facilities (such as roller shutters and epidemic prevention gates). In Embodiment 1, path planning mainly focuses on static or quasi-static obstacles (such as haystacks) within a single work area. However, in cross-area operations, facilities such as roller shutters exhibit time-varying characteristics of "opening" and "closing." A simple spatial potential energy field cannot describe this passageability that evolves over time. Therefore, this embodiment creatively introduces the concept of "V2X spatiotemporal node," mapping physical isolation facilities as dynamic variables in the global potential energy field, and constructing a four-dimensional spatiotemporal potential energy field model that includes the time dimension. At the system architecture level of this technical solution, the pushing robot is equipped with a C-V2X vehicle-mounted unit that supports PC5 direct communication mode. It can perform low-latency bidirectional information interaction with the intelligent roadside unit (RSU) installed on the ranch roller shutter controller. The core of the system lies in establishing a spatiotemporally coupled dynamic potential energy field construction and evolution model. This model no longer regards the roller shutter as a fixed high-repulsive obstacle, but defines it as a potential energy value that changes with time. and facility status To quantify the physical process of dynamically changing "variable impedance nodes," the system defines a spacetime node potential energy function. The function consists of two parts: a static repulsion term based on spatial distance and a time modulation term based on the facility's opening progress. Its mathematical expression is:

[0074] In the formula, Current position of the robot Euclidean distance from the node center; The radius of influence of the potential energy; The basic potential energy gain coefficient; The key time modulation function, whose physical meaning describes the "transparency" of the node, is related to the opening height of the roller shutter door. It exhibits a non-linear inverse proportional relationship; when the roller shutter door is fully closed... Approaching infinity, this node acts as an insurmountable "high potential energy barrier," generating a powerful repulsive force to prevent robot collisions; when the roller shutter door begins to move via V2X commands, its opening height increases with the... The increase, It decays exponentially; when the door is fully open... When the value turns negative, the node instantly flips into a "low potential energy gravity well," or "tunnel mode," which exerts a strong attraction on the robot. This potential energy field flipping mechanism, based on real-time mapping of physical state, ensures that the robot can perceive the possibility of passage from a global perspective, rather than relying solely on local distance sensors. In the specific path planning execution process, this embodiment adopts a spatiotemporal coupling planning strategy based on "arrival time difference" reverse deduction.

[0075] When the robot is still some distance from the target roller shutter door, the system sends an opening request in advance via the V2X link and obtains the mechanical action response time of the roller shutter door. and current communication latency Based on this, the system calculates the absolute time required for the roller shutter door to reach a safe passage state (i.e., the opening height is greater than the robot's height safety threshold). At this point, the robot's speed planning is no longer arbitrary, but subject to strict time constraints; the system defines an ideal arrival speed. Its computational logic follows the spatiotemporal matching formula:

[0076] In the formula, The remaining distance is as follows. The current system time. The physical meaning of this formula, which provides a safety margin of safety within a certain time window, is that the robot must adjust its movement rhythm to ensure it arrives at the door precisely at the exact moment the door opens, avoiding both arriving too early and causing it to stop and wait, and arriving too late and causing the door control system to close prematurely. However, in actual working conditions, due to factors such as slippery ground or load variations, the robot may not be able to maintain such a low speed continuously. (Because motors are inefficient and prone to creeping at extremely low speeds), this embodiment designs an "action chain generation" mechanism. For the waiting phase when the "door is not open," a special "high-damping S-shaped buffer path" is generated. When the system determines that a direct straight approach would lead to premature arrival, the potential energy field planner superimposes a sinusoidal lateral perturbation potential energy field in the preparation area in front of the door. The mathematical model for this perturbation field is:

[0077] in The amplitude coefficient, As a spatial frequency, this term causes the originally straight lines of gravity to twist, inducing the robot to follow a serpentine trajectory. The physical essence of this S-shaped path is to increase the spatial length of the path. To kill time This allows the robot to maintain a high linear velocity. While maintaining momentum and preventing slippage, extending the time to the doorway is similar to an airplane circling and waiting at an airport. This avoids the high-risk "brake-stop-start" sequence on icy roads and eliminates the risk of slippage and loss of control that may occur during the conversion of static friction to kinetic friction. Once the V2X signal confirms that the roller shutter door opening degree has reached the passage threshold, the system immediately triggers the "tunnel straight-through mode." At this time, the aforementioned lateral disturbance potential energy field... The tunnel was instantly removed and replaced by a "low-damping tunnel potential energy zone" constructed within the door frame area. Within this zone, the system artificially adjusts the environmental damping coefficient. Normally, damping fields are used to limit the robot's maximum speed to ensure safety. However, in tunnel mode, to encourage the robot to pass quickly and reduce heat loss caused by the opening of the roller shutter door, the system increases the damping coefficient. Revised to:

[0078] in With an acceleration factor less than 1, and a very strong local gravitational source placed behind the door, the robot experiences a highly uniform net force pointing inwards from the door under this special potential energy field distribution. Furthermore, the drag term is suppressed, resulting in a "sucked-in" acceleration state. To prevent positioning drift during this acceleration, the system locks the heading angle weight at this stage, forcing a specific pushing angle. Returning to zero and retracting the operational posture, traversing narrow spaces with minimal physical cross-section—this transient switch from "S-shaped high-damping buffer" to "straight-line low-damping traversal" is entirely driven by state data in the V2X link, requiring no manual intervention. In addition, the system also introduces an evaluation function based on energy optimality. The energy efficiency ratio is used to monitor the entire spatiotemporal planning process in real time. The evaluation function is defined as:

[0079] The first term represents the power output of the motor, and the second term represents the mechanical impact caused by acceleration and deceleration. The system solves for the minimum value of this functional using the variational method, proving that in cross-regional operations, the S-shaped path that maintains continuous flow can reduce the instantaneous peak current impact by about 30% compared to the start-stop path, and significantly reduce the wear rate of mechanical transmission components.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A path planning method for livestock feeding robot based on AI processing, which is applied to a feeding robot including a multi-source sensing module, an onboard computing unit and an adaptive adjustment push plate; The method includes: The multi-source sensing module is used to collect real-time image data and depth point cloud data of the work area; characterized in that, The method further includes the following steps: Step 1: Process the real-time image data and depth point cloud data using a semantic analysis model to extract the physical attribute features of the forage, which include at least the physical density and the volume of the forage pile. Step 2: Construct a path planning model that includes physical field constraints. The path planning model maps the physical density of forage, the volume of forage pile, and the ground adhesion parameters into motion constraints for the robot. Step 3: Based on the path planning model, the onboard computing unit calculates and generates coupled control commands that include the robot's pose and the pusher deflection angle in real time. Step 4: In response to the coupling control command, the pushing robot moves along the generated hydrodynamic path and dynamically adjusts the deflection angle of the pusher plate, so that the deflection angle of the pusher plate is adjusted in conjunction with the change of the physical property characteristics.

2. The path planning method for a livestock feeding robot based on AI processing according to claim 1, characterized in that, The path planning model adopts a nonlinear model predictive control architecture. The execution steps of the nonlinear model predictive control architecture include: Establish a system prediction model that incorporates the dynamic characteristics of the robot chassis and the nonlinear interaction mechanism between forage and grass; Define the system state vector and control input vector, wherein the control input vector includes at least the drive acceleration, the front wheel steering angle and the push plate angle adjustment angular velocity; Within the prediction time domain, the future state of the robot is iteratively deduced based on the system prediction model. Under the premise of satisfying the physical characteristics constraints of the motor and the ground adhesion constraints, the optimal control input sequence that minimizes the preset multi-objective optimization function is solved.

3. The path planning method for a livestock feeding robot based on AI processing according to claim 2, characterized in that, The mathematical expression of the multi-objective optimization function includes a trajectory tracking error term, an energy consumption cost term, and a control increment penalty term; The trajectory tracking error term is used to characterize the Euclidean distance between the robot's current position and the ideal work line; The energy consumption cost item is calculated based on the motor output driving force, which is derived through the dynamic balance equation. The control increment penalty term is used to limit the rate of change of acceleration, steering angle, and push plate angle.

4. The path planning method for a livestock feeding robot based on AI processing according to claim 2, characterized in that, The system prediction model includes logic for calculating forage pushing resistance, and the calculation of pushing resistance follows the following: in, For the resistance of pushing material, This is the forage deformation resistance coefficient. Forage physical density, Forage pile volume, This represents the current deflection angle of the push plate. The coefficient of kinetic friction between the forage and the ground. The positive pressure generated by the weight of the forage.

5. The path planning method for a livestock feeding robot based on AI processing according to claim 1, characterized in that, The path planning model employs a fuzzy inference system, and the execution steps of the fuzzy inference system include: A fuzzy input model is established to map the physical density of the forage into a fuzzy variable of the forage quantity level, and to map the current feedback value of the pushing robot into a fuzzy variable of the load rate. Construct a fuzzy rule base, which contains multiple logical rules to define the path offset and push plate deflection angle under different combinations of grass level and load rate; The fuzzy inference engine is used to perform calculations on the input fuzzy variables to output a fuzzy control set; the centroid method is used to defuzzify the fuzzy control set to obtain accurate path offset values ​​and push plate angle values.

6. The path planning method for a livestock feeding robot based on AI processing according to claim 5, characterized in that, The final generation of the fluid dynamics path follows a correction logic based on the rate of change of deviation, and its calculation follows the following: in, The target path coordinates, Using the coordinates of the baseline line, For fuzzy output gain coefficient, The path offset values ​​obtained for defuzzification. This is the integral compensation coefficient. For real-time load rate, The target load rate.

7. The path planning method for a livestock feeding robot based on AI processing according to claim 1, characterized in that, The method also includes a correction step based on optical flow anti-slip: Ground texture features are collected using a downward-looking visual sensor, and the average pixel movement speed of the ground is calculated using the dense optical flow method. The average pixel movement speed is converted into physical observation speed by combining the camera intrinsic parameter matrix; The difference between the physically observed velocity and the theoretical velocity fed back by the robot wheel speedometer is calculated to obtain the actual lateral slip velocity vector; An anti-slip compensation force vector, which is opposite in direction to the actual lateral slip velocity vector, is superimposed on the path planning model.

8. The path planning method for a livestock feeding robot based on AI processing according to claim 7, characterized in that, The step of generating coupled control commands that include robot pose and pusher deflection angle also includes generating crab approach control commands: Based on the anti-slip compensation force vector and the resultant force vector of the path planning model, the crab-shaped compensation angle is calculated. The control of the robot's chassis yaw angle deflects the crab-shaped compensation angle, so that the robot's front direction forms an angle with the actual speed direction, and the longitudinal component of the wheel rolling friction force is used to counteract lateral slippage.

9. The path planning method for a livestock feeding robot based on AI processing according to claim 1, characterized in that, The method also includes a global planning step based on V2X spatiotemporal nodes: The on / off status and action response time of the pasture's physical barriers are obtained through the V2X communication link; Establish a spatiotemporal node potential energy function for the physical partition facility in the global map; The spatiotemporal node potential energy function includes a static repulsive term based on spatial distance and a time modulation term based on the facility's activation progress. The value of the time modulation term decreases as the facility opening height increases.

10. The path planning method for a livestock feeding robot based on AI processing according to claim 9, characterized in that, The global planning step based on V2X spatiotemporal nodes also includes action chain generation logic: Based on the time difference of arrival reverse reasoning strategy, the ideal arrival speed of the robot to the physical partition facility is calculated; When the facility is not fully open and the robot will arrive earlier than the opening time if it travels in a straight line, a high-damping S-shaped buffer path is generated to control the robot to travel along the S-shaped trajectory to consume time. When the facility's opening level reaches the passage threshold, a low-damping tunnel straight path is generated, and the environmental damping coefficient is adjusted to control the robot to accelerate through.

Images