A Systematic Review of Modeling and Control Approaches for Path Tracking in Unmanned Agricultural Ground Vehicles

Abstract

With the advancement of precision agriculture, the autonomous navigation of unmanned agricultural ground vehicles (UAGVs) has emerged as a critical research topic. As a fundamental component of autonomous navigation, path-tracking control is essential for ensuring the accurate and stable operation of UAGVs. However, achieving high-precision and robust tracking in agricultural environments remains challenging due to unstructured terrain, variable wheel slip, and complex dynamic disturbances. This review provides a structured and comprehensive survey of modeling and control methodologies for UAGVs, with particular emphasis on control-theoretic formulations and their applicability across diverse agricultural scenarios. In contrast to prior reviews, the modeling approaches are systematically classified into geometric, kinematic, and dynamic models, including extended formulations that incorporate wheel slip and external disturbances. Furthermore, this paper systematically reviews commonly adopted path-tracking strategies for UAGVs, including proportional–integral–derivative (PID) control, pure pursuit (PP), Stanley control, sliding mode control (SMC), model predictive control (MPC), and learning-based approaches. Emphasis is placed on their theoretical underpinnings, tracking accuracy, adaptability to unstructured field environments, and computational efficiency. In addition, several key technical challenges are identified, such as terrain-adaptive vehicle modeling, slip compensation mechanisms, real-time implementation under hardware constraints, and the cooperative control of multiple UAGVs operating in dynamic agricultural scenarios. By presenting a detailed review from a control-centric perspective, this study aims to serve as a valuable reference for researchers and practitioners developing intelligent agricultural vehicle systems.

1. Introduction

The integration of next-generation information technologies into agriculture is accelerating the transformation of modern agricultural systems towards networked, digitalized, and intelligent paradigms. This transformation, driven by the agricultural digital revolution, has fostered the emergence of smart agriculture as a new production mode [1,2]. Contemporary agricultural development emphasizes improving resource-use efficiency and safeguarding agroecological systems [3,4], thereby imposing higher requirements on agricultural machinery. The shift toward intelligent and digital equipment has become a prevailing trend in machinery design and development [5,6]. Among these technological advancements, autonomous navigation has become a central focus in the modernization of agricultural vehicles [7,8]. A complete autonomous navigation system typically comprises four key modules: environmental perception [9,10,11,12], path planning [13,14], path-tracking, and operation control [15,16,17,18,19]. Among these, the path-tracking controller plays a critical role in ensuring the operational safety and precision of unmanned agricultural ground vehicles (UAGVs). Consequently, recent research efforts have increasingly focused on vehicle modeling and the development of robust tracking control strategies to enable accurate trajectory following in complex and unstructured field environments [20].

In contrast to the relatively structured and predictable conditions found in industrial environments, agricultural fields are inherently unstructured, dynamic, and highly variable [21,22,23,24]. Challenges such as uneven terrain, fluctuating soil resistance, and low-adhesion surfaces introduce substantial uncertainties that hinder accurate system modeling and robust control performance in UAGV systems [25,26]. When directly applied to these conditions, conventional control strategies often suffer from degraded accuracy, poor disturbance rejection, and limited generalization capabilities. Therefore, the development of high-fidelity mathematical models tailored to the nonlinear and time-varying characteristics of agricultural environments—together with the design of path-tracking controllers that ensure adaptability and robustness under external disturbances—is essential for realizing safe and efficient autonomous navigation in UAGVs [27].

Accurate vehicle modeling constitutes the foundation of path-tracking controller design, as the fidelity of the model directly influences the tracking accuracy and control robustness [28]. Existing modeling approaches for UAGVs can be broadly classified into geometric [29], kinematic [30], and dynamic models [31], each featuring distinct assumptions and applicability. Geometric models describe spatial relationships—such as heading angle, turning radius, and vehicle pose—under idealized assumptions that ignore system mass, tire forces, and dynamic effects [32]. These models support classical tracking algorithms including PID, PP, and Stanley methods. Due to their simplicity and low computational demands, geometric models are particularly applicable to low-speed operations in structured agricultural environments. Kinematic models, in contrast, capture the relationship between control inputs (e.g., linear and angular velocities) and vehicle pose evolution over time. Their low dimensionality and clearly defined state variables render them suitable for advanced control schemes such as SMC and MPC. For enhanced robustness, extended kinematic models incorporate external disturbances arising from wheel slip, terrain irregularities, or payload variations, and are often integrated with compensation techniques. Dynamic models provide higher-fidelity representations by accounting for the vehicle’s response to external forces, including inertial dynamics, tire–ground interactions, and side slip. Typically derived from Newton–Euler mechanics, these models are essential for high-speed navigation, variable payloads, and operations on unstructured or inclined terrain. However, their practical application is often limited by structural complexity, sensitivity to parameter uncertainties, and computational overhead. To address these challenges, various adaptive strategies—such as disturbance observers, online parameter identification, and learning-based modeling using real-time data—have been proposed to enhance robustness and maintain control performance under uncertain field conditions.

In summary, achieving high-precision path-tracking control for UAGVs in complex and unstructured agricultural environments requires the coordinated integration of vehicle modeling and control strategy design to ensure robustness and adaptability. Although substantial progress has been achieved, prior reviews often lack a thorough analysis from a control-theoretic perspective. To bridge this gap, this paper presents a structured and in-depth survey of modeling methodologies and path-tracking control strategies for UAGVs, with a particular emphasis on their suitability across diverse agricultural scenarios. Furthermore, it highlights key technical challenges—such as terrain-adaptive modeling, slip compensation mechanisms, and real-time implementation under computational constraints—and outlines promising directions for future research. This review aims to provide a focused, control-centric synthesis to assist researchers and practitioners in developing intelligent, robust, and efficient UAGV systems for field operations. Figure 1 presents the overall structure of this review, linking modeling strategies, control algorithms, sensor integration, application scenarios, and future research directions.

The remainder of this paper is organized as follows. Section 2 outlines the review methodology. Section 3 discusses common modeling approaches for unmanned agricultural ground vehicles (UAGVs), including geometric, kinematic, and dynamic models, as well as extended formulations designed to address typical uncertainties in agricultural environments. Section 4 provides a comprehensive overview of path-tracking control algorithms for UAGVs. Section 5 examines the integration of proximal sensors with UAGVs, while Section 6 reviews the deployment of control strategies across different agricultural scenarios. Section 7 presents the results of the review. Finally, Section 8 concludes the paper and highlights promising directions for future research.

2. Review Methodology

This review adheres to the PRISMA 2020 guidelines [33] to ensure transparency, reproducibility, and methodological rigor. The methodology encompasses the literature search, eligibility screening, and analytical procedures. Relevant publications were identified from leading academic databases, including Web of Science, Scopus, IEEE Xplore, and ScienceDirect. The search was limited to English-language peer-reviewed journal articles, conference papers, and technical reports published between 2015 and June 2025, thereby capturing recent advances in the field.

A set of targeted keywords was employed to retrieve studies related to the modeling and control of UAGVs. These included terms such as “unmanned agricultural systems,” “agricultural navigation,” “vehicle modeling,” “path tracking,” “pure pursuit control,” “Stanley control,” “PID control,” “sliding mode control,” “model predictive control,” and “learning-based control.” Boolean operators (AND, OR, NOT) were used to construct logical queries and refine search results.

2.1. Eligibility Criteria

To ensure methodological consistency, a set of predefined inclusion and exclusion criteria was applied during the screening process, as detailed in

Table 1. Summary of inclusion and exclusion criteria.

Inclusion Criteria	Exclusion Criteria
Modeling and control of UAGVs	Studies on UAVs or non-agricultural platforms
Field-tested or validated in agricultural context	Pure simulation without practical relevance
Focus on control algorithms	Studies limited to perception, mapping, or planning
Published in peer-reviewed SCI/EI/Scopus venues	Non-peer-reviewed or grey literature
Available in English full text	Inaccessible or non-English publications

.

2.2. Screening and Selection Process

An initial query across major academic databases—including Web of Science, Scopus, IEEE Xplore, and ScienceDirect—identified 198 records related to UAGV modeling and path-tracking control. After removing duplicates, 167 unique records remained for screening. The screening process was conducted in two stages. First, 35 records were excluded based on abstract review due to topic irrelevance (e.g., UAVs, general autonomous vehicles, or non-agricultural platforms).

Second, full-text screening excluded 25 records that lacked practical relevance, such as studies limited to theoretical simulation without agricultural context or real-world validation. Ultimately, 107 articles met the inclusion criteria and were retained for systematic review. These studies focus on UAGV modeling and control strategies and were published in peer-reviewed journals or conferences indexed by SCI, EI, or Scopus. In addition, 33 other references were cited for contextual or background purposes—such as the concepts of digital agriculture, machinery components, and autonomous system architecture—but were not included in the systematic review. Figure 2 illustrates the PRISMA 2020 flow diagram summarizing the study selection and inclusion process.

2.3. Descriptive Statistics

Figure 3a illustrates a steady increase in publications related to UAGV modeling and control from 2015 to 2024, with a marked acceleration observed after 2021. This trend underscores the growing academic and industrial interest in intelligent agricultural machinery, fueled by the rising demand for automation and digital transformation in modern farming systems.

Figure 3b presents the distribution of control strategies adopted across the surveyed literature. Among them, MPC and SMC have emerged as dominant approaches, attributed to their strong robustness and adaptability to field uncertainties. Classical methods such as PID control and PP remain widely utilized due to their algorithmic simplicity and practical ease of deployment. In parallel, learning-based control strategies have gained momentum in recent years, reflecting a broader shift towards data-driven and adaptive control paradigms in agricultural automation.

Figure 4 illustrates the geographic distribution of first-author affiliations across the reviewed studies. A substantial proportion of contributions originate from Asia, with China, India, and South Korea emerging as leading contributors. This concentration is likely driven by large agricultural populations, increasing demand for mechanization, and sustained investment in digital agricultural technologies. North America and Europe also demonstrate significant research activity, supported by mature agricultural systems and well-established academic institutions. In contrast, Africa, South America, and Oceania account for relatively fewer publications, which may reflect limitations in research funding, lower mechanization levels, and nascent digital agriculture infrastructures. These regional disparities underscore the importance of fostering international collaboration to accelerate the global development and adoption of UAGV technologies.

3. Modeling Strategies for Unmanned Agricultural Ground Vehicles

Accurate modeling forms the foundation of path planning and tracking control for UAGVs [34]. High-fidelity models are essential for capturing the constraints, nonlinearities, and external disturbances inherent to real-world UAGV operations [35,36]. Given the diversity of UAGV platforms and the complexity of agricultural terrains—ranging from grasslands and dry fields to paddy fields and sloped environments—selecting appropriate modeling strategies is critical for achieving precise trajectory tracking and ensuring robust control performance.

This section presents representative modeling approaches commonly employed in UAGV control design, including geometric, kinematic, dynamic, and extended models that incorporate environmental uncertainties. Each modeling framework offers distinct advantages and limitations, which must be carefully considered based on specific control objectives and operational scenarios.

3.1. Geometric Modeling

As illustrated in Figure 5, geometric modeling characterizes the spatial relationship between a vehicle’s wheelbase and its turning trajectory, emphasizing curvature during steering maneuvers. In contrast to kinematic and dynamic models, geometric models neglect velocity and acceleration dynamics, focusing solely on spatial configuration and path geometry. This abstraction is particularly useful for analyzing how the physical layout of the vehicle affects its maneuverability.

Among geometric modeling approaches applied to UAGVs, the Ackermann steering geometry is the most prevalent. It assumes that all wheels instantaneously rotate about a common instantaneous center of rotation (ICR) during turning [37,38]. Based on this assumption, the steering angle δ is related to the turning radius R and wheelbase L through the following relationship:

$$\delta ={tan}^{-1}\left({\displaystyle \frac{L}{R}}\right)$$

(1)

Notably, the geometric modeling approach forms a critical theoretical foundation for widely adopted path-tracking controllers such as PP and Stanley control. Implementation details and comparative analyses of these two controllers are presented in subsequent sections.

3.2. Kinematic Modeling

Kinematic models are extensively used in agricultural robotics owing to their structural simplicity and effectiveness in low-speed, structured environments [39]. They offer a favorable trade-off between modeling fidelity and computational efficiency, which makes them particularly suitable for UAGVs operating in flat terrains or along predefined crop rows. Nevertheless, their accuracy tends to decline in unstructured or uneven agricultural fields, where increased terrain variability and external disturbances pose significant modeling and control challenges.

3.2.1. Kinematic Modeling of Tracked Chassis

As illustrated in Figure 6a, tracked chassis configurations are commonly used in differential drive-driven UAGVs, such as tracked harvesters, tracked tractors, and lawn mowers. These vehicles typically achieve directional control by independently varying the velocities of the left and right tracks or wheels [40]. Their motion can therefore be modeled using the basic unicycle kinematic model, which relates the vehicle’s linear and angular velocities directly to the speeds of the left and right tracks:

$$v={\displaystyle \frac{v_r+v_l}{2}},\omega ={\displaystyle \frac{v_r-v_l}{B}}$$

(2)

where v_r and v_l denote the velocities of the right and left tracks, respectively, and B represents the separation distance between the tracks.

As depicted in Figure 6b, a widely adopted kinematic representation for tracked vehicles is the ideal unicycle model, valued for its simplicity and suitability in control design. This model approximates the vehicle’s motion as that of a point mass, characterized solely by a forward velocity v and an angular velocity ω. The corresponding state equations are

$$\left\{\begin{array}{c}\dot{x}=vcos\theta \hfill \\ \dot{y}=vsin\theta \hfill \\ \dot{\theta }=\omega \hfill \end{array}\right.$$

(3)

where (x, y) represents the vehicle’s positional coordinates and θ denotes its heading angle.

To more accurately capture the skid-steering behavior of tracked UAGVs under realistic agricultural conditions, recent studies [41,42,43] have proposed extensions to the ideal kinematic model by incorporating slip-related components. These enhanced formulations aim to mitigate the loss of modeling fidelity caused by terrain irregularities and complex wheel–soil interactions, which frequently violate the assumptions underlying conventional kinematic models. The resulting extended kinematic model can be expressed as

$$\begin{array}{c}\dot{x}={\displaystyle \frac{r}{2}}[{\omega }_L(1-i_L)+{\omega }_R(1-i_R)][cos\theta -sin\theta tan\alpha ]\hfill \\ \dot{y}={\displaystyle \frac{r}{2}}[{\omega }_L(1-i_L)+{\omega }_R(1-i_R)][sin\theta +cos\theta tan\alpha ]\hfill \\ \dot{\theta }={\displaystyle \frac{r}{B}}[{\omega }_L(1-i_L)-{\omega }_R(1-i_R)]\hfill \end{array}$$

(4)

where α is the slip angle reflecting lateral displacement, r is the radius of the track-driving sprockets, ω_L and ω_R denote the angular velocities of the left and right tracks, respectively, and i_L and i_R represent their corresponding slip ratios.

This model explicitly incorporates environmental disturbances—such as lateral slippage and asymmetric traction loss—through parameters like slip angle and slip ratio. By accounting for these non-ideal effects, the model significantly improves adaptability to complex agricultural terrains. Consequently, it enhances the robustness of trajectory planning and control in real-world field conditions.

3.2.2. Kinematic Modeling of Ackermann Chassis

For UAGVs with Ackermann steering—such as tractors, rice transplanters, and sprayers—the bicycle model offers a suitable kinematic representation [44]. As illustrated in Figure 7, this simplified model reduces a standard four-wheel chassis to an equivalent two-wheel abstraction, where the front wheel is responsible for directional steering and the rear wheel provides propulsion. The ideal kinematic equations for the bicycle model are given by

$$\left\{\begin{array}{c}\dot{x}=vcos\theta \hfill \\ \dot{y}=vsin\theta \hfill \\ \dot{\theta }={\displaystyle \frac{vtan\delta }{L}}\hfill \end{array}\right.$$

(5)

where v represents the forward velocity, δ denotes the steering angle of the front wheel, and L is the vehicle’s wheelbase.

The ideal bicycle kinematic models rely on simplified assumptions—such as the absence of wheel slip—which limit their applicability in agricultural environments [45]. UAGVs frequently operate on challenging terrain, including wet grasslands, muddy fields, and loose or uneven surfaces, where both longitudinal and lateral slip are inevitable. These slip-induced effects degrade model accuracy, impair path-tracking performance, and may destabilize the control system. To overcome these limitations, extended kinematic models incorporating slip parameters have been proposed [46]. By explicitly modeling slip dynamics, these formulations offer enhanced motion fidelity and robustness in unstructured field conditions, thereby improving the reliability of trajectory tracking.

For Ackermann steering UAGVs, the extended kinematic model that incorporates slip velocities and slip angles is given as follows [47]:

$$\left\{\begin{array}{c}\dot{x}=(v-v_l^s)cos\theta -v_s^{s}sin\theta \hfill \\ \dot{y}=(v-v_l^s)sin\theta +v_s^{s}cos\theta \hfill \\ \dot{\theta }={\displaystyle \frac{(v-v_l^s)}{L}}tan(\delta +{\delta }_s)+{\displaystyle \frac{v_s^s}{L}}\hfill \end{array}\right.$$

(6)

where $v_l^s$ denotes the longitudinal slip velocity, $v_s^s$ is the lateral slip velocity, and δ_s represents the additional slip-induced steering angle caused by lateral sliding. By explicitly incorporating these slip parameters, the extended model compensates effectively for deviations induced by wheel slip, thereby enhancing the trajectory tracking precision and improving the overall control system stability.

In the context of UAGV path-tracking research, extended kinematic models have gained increasing attention due to their capability to capture slip-induced dynamics. By explicitly modeling both longitudinal and lateral slip, these formulations facilitate the design of advanced controllers and adaptive algorithms that estimate and compensate for slip effects in real time. This improves trajectory accuracy and enhances system robustness under varying terrain conditions commonly encountered in agricultural operations.

3.3. Dynamic Modeling

Dynamic modeling has attracted growing interest due to its ability to characterize the complex motion behavior of UAGVs under realistic field conditions. In contrast to kinematic models that primarily describe geometric relationships, dynamic models account for vehicle mass, inertial effects, acceleration, friction, and tire–terrain interaction forces, thereby capturing the physical response of the system to control inputs. These models are particularly valuable in scenarios involving unstructured terrain, varying payloads, and operations requiring high-precision tracking, where slip, vibration, and terrain-induced disturbances are significant. Although computationally intensive, the implementation of dynamic models in UAGV systems has been increasingly enabled by advancements in real-time computing and model identification techniques, supporting more accurate and robust control in complex agricultural environments.

3.3.1. Dynamic Modeling of Tracked Chassis

Tracked vehicles, driven independently by left and right tracks, offer strong off-road mobility due to their large ground contact area. However, their motion involves complex dynamics influenced by soil–track interaction, rolling resistance, and slip. A simplified dynamic model capturing longitudinal, lateral, and yaw motions is shown in Figure 8.

$$\left\{\begin{array}{c}m\ddot{x}=F_{ql}+F_{qr}-F_{rl}-F_{rr}\hfill \\ m\ddot{y}=4f_y{d}_s\hfill \\ I_z\ddot{\theta }=(F_{ql}-F_{rl}){\displaystyle \frac{B}{2}}-(F_{qr}-F_{rr}){\displaystyle \frac{B}{2}}-2f_y(4l^2-d_s^2)\hfill \end{array}\right.$$

(7)

where m represents the vehicle’s mass, I_z is the yaw moment of inertia, B denotes the track width, and F_ql, F_qr and F_rl, F_rr correspond, respectively, to the driving forces and rolling resistances acting on the left and right tracks. The parameter f_y represents lateral forces, while d_s is the lateral offset between the vehicle’s centerline and the track contact patches, and l characterizes the longitudinal geometry of the chassis.

Despite their practical utility, simplified dynamic models often neglect key real-world effects—such as track slip, lateral tire-–terrain interaction, and external disturbances—that are frequently encountered during agricultural field operations. These omissions may result in considerable discrepancies between the predicted and actual vehicle behavior, thereby degrading path-tracking accuracy and potentially compromising overall system stability. To overcome these limitations, extended dynamic models have been proposed, which explicitly incorporate slip dynamics and disturbance terms into the equations of motion. Such enhancements improve model fidelity and control robustness under complex and variable agricultural conditions [48]. The general form of an extended dynamic model is expressed as

$$\left\{\begin{array}{c}m\ddot{x}=F_{ql}+F_{qr}-F_{rl}-F_{rr}+d_x\hfill \\ m\ddot{y}=4f_y{d}_s+d_y\hfill \\ I_z\ddot{\theta }=(F_{ql}-F_{rl}){\displaystyle \frac{B}{2}}-(F_{qr}-F_{rr}){\displaystyle \frac{B}{2}}-2f_y(4l^2-d_s^2)+d_{\theta }\hfill \end{array}\right.$$

(8)

where m denotes the vehicle mass, I_z is the yaw moment of inertia about the vertical axis, F_ql, F_qr and F_rl, F_rr, respectively, represent the longitudinal driving and resistive forces acting on the left and right wheels. Additionally, f_y denotes the lateral tire–terrain interaction force, d_s is the lateral offset of tire–ground contact, B is the vehicle track width, and l characterizes the longitudinal geometry of the chassis. Disturbances from uneven terrain or external factors are captured by the additional disturbance terms d_x, d_y, and d_θ.

3.3.2. Dynamic Modeling of Ackermann Chassis

For UAGVs equipped with Ackermann steering—such as tractors, sprayers, and other wheeled agricultural machinery—dynamic modeling must account for the coupled effects of front-wheel steering, rear-wheel drive, mass distribution, and vehicle orientation. In contrast to kinematic models that neglect inertial and resistive forces, dynamic models explicitly incorporate vehicle mass, rotational inertia, tire–terrain interaction, and external disturbances [49]. These considerations are essential for accurately representing the vehicle’s motion response under complex field conditions, particularly when operating at higher speeds or over uneven terrain.

A representative dynamic model of an Ackermann-steered vehicle, as illustrated in Figure 9, captures the vehicle’s longitudinal, lateral, and yaw dynamics. The model can be formulated as follows:

$$\left\{\begin{array}{c}m\ddot{x}=m\dot{y}\dot{\theta }+2(F_{lf}cos\delta -F_{cf}sin\delta )+2F_{lr}\hfill \\ m\ddot{y}=-m\dot{x}\dot{\theta }+2(F_{lf}sin\delta +F_{cf}cos\delta )+2F_{cr}\hfill \\ I_Z\ddot{\theta }=2l_f(F_{lf}sin\delta +F_{cf}cos\delta )-2l_r{F}_{cr}\hfill \end{array}\right.$$

(9)

where F_lf and F_lr denote longitudinal tire forces acting on the front and rear of the chassis, respectively; F_cf and F_cr indicate corresponding lateral tire forces; l_f and l_r represent the distances from the vehicle’s center of gravity to the front and rear tire–ground contact points, respectively; and δ is the effective slip angle incorporating lateral slip effects.

While simplified dynamic models provide a foundational framework, they often neglect critical vehicle–terrain interaction effects, including lateral and longitudinal slip, steering-induced slip angles, and various unmodeled dynamics that substantially influence UAGV behavior. These factors are particularly significant in agricultural environments—such as muddy, slippery, or uneven terrain—where they can degrade path-tracking accuracy and compromise vehicle stability. To overcome these limitations, extended dynamic models have been proposed that explicitly incorporate slip phenomena, lateral tire forces, and external disturbance torques, thereby offering a more accurate and robust representation of Ackermann-steered vehicles operating under realistic field conditions [50]. The corresponding extended dynamic model is formulated as

$$\left\{\begin{array}{c}m\ddot{x}=m\dot{y}\dot{\theta }+2(F_{lf}cos\delta -F_{cf}sin\delta )+2F_{lr}+d_x\hfill \\ m\ddot{y}=-m\dot{x}\dot{\theta }+2(F_{lf}sin\delta +F_{cf}cos\delta )+2F_{cr}+d_y\hfill \\ I_Z\ddot{\theta }=2l_f(F_{lf}sin\delta +F_{cf}cos\delta )-2l_r{F}_{cr}+d_{\theta }\hfill \end{array}\right.$$

(10)

where F_lf and F_lr represent the longitudinal tire forces at the front and rear wheels, respectively, while F_cf and F_cr correspond to the lateral tire forces. The parameters l_f and l_r denote the distances from the vehicle’s center of mass to the front and rear tire–ground contact points, respectively. The variable δ represents the effective front-wheel steering angle, incorporating variations induced by slip. The terms d_x, d_y, and d_θ account for external disturbances and terrain-induced variations that affect vehicle motion.

3.4. Summary and Comparative Analysis

Geometric, kinematic, and dynamic models represent the primary modeling paradigms for the navigation and control of UAGVs. Each category entails varying degrees of complexity, computational burden, and fidelity in capturing real-world dynamics. The selection of an appropriate modeling approach is highly dependent on terrain characteristics, operational objectives, and the required control accuracy.

In agricultural applications, model selection becomes particularly critical due to the unstructured, variable nature of field environments. Geometric models are suitable for path planning in structured settings (e.g., flat farmland or greenhouses), where idealized assumptions—such as no slip—are approximately valid. These models are computationally efficient but exhibit limited performance in low-traction or dynamically changing conditions [51]. Kinematic models provide a practical compromise between model simplicity and physical realism, and are frequently employed in tasks such as row-following or turning in moderately structured environments like orchards. Although effective at low speeds, their omission of wheel–soil interaction and slip effects restricts their utility in more challenging terrain [52,53]. Dynamic models are essential for high-precision field operations (e.g., automated tillage, variable-rate spraying) conducted under uncertain or highly dynamic conditions. By explicitly accounting for vehicle inertia, slip behavior, terrain interaction, and payload variation, these models enable robust trajectory tracking in complex agricultural scenarios [54,55,56]. Although more computationally intensive, recent advances in onboard processing and sensing technologies have increasingly facilitated their practical deployment.

A comparative overview of these modeling strategies is presented in

Table 2. Comparison of geometric, kinematic, and dynamic models.

Model Type	Complexity	Slip Handling	Application Scenarios
Geometric model	Low	Not considered	Low-speed navigation or PP/Stanley method
Ideal kinematic model	Medium	Not considered	Flat terrain, low-speed navigation
Extended kinematic model	Medium	Longitudinal and lateral slip	Uneven, slippery or unknown terrains
Ideal dynamic model	High	Not considered	Structured dynamic conditions, known load and terrain
Extended dynamic model	High	Explicit modeling of slip and lateral tire forces	Complex agricultural fields, variable loads

, summarizing their respective assumptions, performance trade-offs, and application suitability across diverse agricultural contexts.

4. Path-Tracking Control Strategies

4.1. PID Control

PID control remains one of the most commonly employed techniques in agricultural UAGVs, owing to its structural simplicity and minimal computational requirements [57]. It is particularly effective in structured or semi-structured environments—such as row-crop fields, greenhouses, and flat farmlands—where external disturbances and terrain variability are relatively limited. In path-tracking applications, the classical PID control law is typically formulated as

$$u\left(t\right)=K_{p}e\left(t\right)+K_i{\int }_0^{t}e\left(\tau \right)d\tau +K_d{\displaystyle \frac{de\left(t\right)}{dt}}$$

(11)

where e(t) denotes the real-time tracking error, and K_p, K_i, and K_d are the proportional, integral, and derivative gains, respectively. The proportional term provides immediate correction based on the current error; the integral term accumulates past errors to eliminate steady-state offset; and the derivative term anticipates future error trends, enhancing system damping and reducing overshoot (Figure 10).

While PID-based control remains attractive due to its structural simplicity and ease of implementation, it exhibits limited adaptability under significantly varying operating conditions. Parameter tuning is often conducted empirically and may yield suboptimal performance in dynamic or uncertain field scenarios. To enhance PID control under practical agricultural conditions, recent studies have explored hybrid strategies that integrate PID with fuzzy logic, adaptive gain adjustment, or preview control mechanisms. For instance, in intelligent weeding systems, fuzzy adaptive PID controllers dynamically adjusted gains based on real-time steering deviations, thereby improving responsiveness and noise robustness in densely planted row environments [58]. In tractor navigation tasks, a model-free predictive PID framework demonstrated robust performance across heterogeneous terrain profiles and during complex turning maneuvers [59].

Preview-based PID designs have further improved heading stability in rice transplanters and other UAGVs by replacing the proportional term with a predicted heading deviation and tuning control gains using fuzzy logic [60,61]. Self-tuning fuzzy PID controllers have also been implemented in autonomous tractors, effectively reducing heading error and enhancing adaptability under varying load conditions and traction scenarios [62,63].

Table 3. Summary of improved PID-based strategies for agricultural path tracking.

PID Type	Improved Method	Performance Highlights	References
Fuzzy-tuned PID	Fuzzy logic adjusts PID gains based on error trends	Improved responsiveness; better path adherence in turns	[58,60,61]
Model-free adaptive predictive PID	Prediction-based adaptive control structure embedding PID core	Maintains robustness across terrain and path changes for tractors	[59]
Adaptive PID	Adaptive fuzzy logic for continuous online gain tuning	Improved heading stability and positioning in autonomous tractors	[62,63]

summarizes representative enhancements to PID-based control strategies for agricultural path tracking. By incorporating fuzzy logic, adaptive tuning, preview modeling, and hierarchical control structures, these methods improve the flexibility and accuracy of PID controllers across a wide range of field conditions, thereby mitigating the inherent limitations of conventional PID schemes.

4.2. Pure Pursuit Method

The PP algorithm is among the most widely employed geometric path-tracking strategies in autonomous agricultural navigation, owing to its conceptual simplicity, low computational overhead, and reliable performance at low to moderate speeds. It determines the steering command by constructing a circular arc that connects the vehicle’s current position to a target point located at a predefined look-ahead distance along the reference trajectory. This arc defines the instantaneous turning radius required for the vehicle to converge to the path, making the approach particularly effective for UAGVs operating in structured or semi-structured environments (Figure 11).

Given a look-ahead distance l_d and a heading error angle α between the vehicle’s current orientation and the target point, the required turning radius R and the corresponding front-wheel steering angle δ_f can be computed using the following geometric relationships:

$$R={\displaystyle \frac{l_d}{2sin\alpha }},{\delta }_f=arctan\left({\displaystyle \frac{2lsin\alpha }{l_d}}\right)$$

(12)

The look-ahead distance (l_d) plays a critical role in determining the performance of PP-based controllers. A shorter l_d enhances controller responsiveness but may induce oscillatory behavior, particularly in high-curvature segments. In contrast, a longer l_d produces smoother trajectories but can reduce tracking accuracy, especially when negotiating sharp turns or irregular paths. To address this trade-off, recent studies have proposed adaptive mechanisms that dynamically adjust l_d based on real-time environmental and trajectory features.

For instance, fuzzy logic-based tuning strategies that adjust l_d according to lateral deviation and path curvature have demonstrated reductions in lateral error as low as 1.8 cm under field conditions [64]. In tracked UAGVs employing brake steering, curvature-aware l_d adaptation achieved an average tracking error reduction of 15.6% [65]. Additional enhancements include incorporating heading error rate and curvature segmentation to accelerate convergence and improve control stability [66]. Furthermore, hybrid strategies integrating PP with MPC-based preview control have been developed, wherein l_d is continuously varied in response to terrain conditions and path smoothness, resulting in improved tracking precision and smoother trajectory execution in unstructured agricultural environments [67].

Table 4. Summary of improved PP-based strategies for agricultural path tracking.

PP Type	Improved Method	Performance Highlights	References
Optimal goal-point PP	Online optimization of goal-point to minimize heading and lateral error	Improved stability and convergence speed in variable curvature paths	[65,66]
Fuzzy PP	Fuzzy logic adjusts look-ahead distance (ld) based on real-time error and curvature	Reduced lateral tracking error in both simulation and field environments	[64]
MPC-based PP	Preview distance modulated via MPC to adaptively balance curvature sensitivity	Reduced overall tracking error and improved path smoothness	[67]

summarizes key advancements in PP-based tracking methods tailored for agricultural navigation. These approaches adapt the look-ahead distance dynamically based on factors such as curvature, error trends, and obstacle proximity. By accounting for the variability of real field conditions, they improve tracking accuracy, responsiveness, and path-following robustness in diverse agricultural terrains.

4.3. Stanley Control Method

Originally developed for autonomous driving in high-speed racing scenarios, the Stanley control method has been widely adopted in UAGV navigation due to its geometric simplicity and robust closed-loop feedback structure [68]. Unlike the PP algorithm, which computes control actions based on a look-ahead point, the Stanley controller directly utilizes the real-time feedback of both lateral deviation y_e and heading error θ_e between the vehicle and the reference trajectory. Its steering control law is formulated as

$${\delta }_f={\theta }_e+arcsin\left({\displaystyle \frac{k{y}_e}{v}}\right)$$

(13)

where v is the vehicle’s forward speed and k is a control gain. The first term aligns the vehicle’s heading with the trajectory, while the second term provides lateral correction that is inversely scaled with speed. This design improves stability at higher velocities and eliminates the need for manually tuning a look-ahead distance, which is typically required in preview-based algorithms (Figure 12).

To enhance the adaptability of the Stanley controller under diverse agricultural operating conditions, fuzzy logic has been employed to dynamically adjust the gain parameter k based on real-time tracking errors. This approach improves control responsiveness and stability during both straight-line navigation and turning maneuvers. To further optimize gain tuning, particle swarm optimization (PSO) has been integrated with fuzzy logic, enabling adaptive gain scheduling under varying speeds and actuator constraints. Simulation results have demonstrated that this method can reduce lateral error by an order of magnitude [69]. Additionally, genetic algorithm (GA)-based tuning strategies have been investigated, producing enhanced Stanley controllers capable of reducing root mean square tracking error by up to 48.6% in representative agricultural scenarios such as U-turns and Ω-shaped paths [70].

Table 5. Summary of improved Stanley-based strategies for agricultural path tracking.

Stanley Type	Improved Method	Performance Highlights	References
PSO–Stanley	PSO optimizes controller parameters to adapt gain	Ten-fold reduction in lateral error under variable speeds and constraints	[69]
GA–Optimized Stanley	Genetic algorithm used to tune gain for different path	Improved control precision maintained under diverse terrain and route conditions	[70]

summarizes recent enhancements to the Stanley controller tailored for agricultural path-tracking applications. These improvements—primarily through the integration of fuzzy logic and evolutionary optimization algorithms—enable the adaptive tuning of control gains in response to variations in path curvature, vehicle speed, and actuator limitations. As a result, significant improvements in tracking accuracy have been achieved under realistic field conditions. Nevertheless, ensuring control stability in scenarios involving abrupt curvature transitions or operating on low-adhesion surfaces remains a persistent challenge requiring further investigation.

4.4. Sliding Mode Control

SMC has been widely adopted in UAGV applications due to its inherent robustness against model uncertainties, external disturbances, and system nonlinearities [71,72,73]. Conventional SMC designs, typically based on simplified kinematic models, have demonstrated reliable trajectory convergence in structured environments such as flat farmlands. However, their performance often deteriorates in complex agricultural scenarios—characterized by uneven terrain, variable soil–vehicle interactions, and frequent wheel slip—where traditional SMC exhibits reduced tracking accuracy and pronounced control chattering [74].

To improve adaptability in complex and variable agricultural environments, adaptive sliding mode control (ASMC) strategies have been proposed. These approaches adjust control gains online to accommodate time-varying or uncertain system parameters—such as cornering stiffness, soil resistance, and payload fluctuations—in real time [75,76]. While ASMC enhances robustness and responsiveness under diverse conditions, it also introduces sensitivity to the tuning of adaptation rates, which must be carefully designed to prevent instability, overshoot, or degraded transient performance [77].

In addition to adaptive mechanisms, fuzzy logic has been widely incorporated into SMC frameworks to suppress high-frequency switching and reduce control chattering [78,79]. Fuzzy inference systems are often employed to adjust the switching gain dynamically or to construct continuous sliding surfaces, thereby improving transient response and minimizing actuator stress. These improvements are particularly beneficial for agricultural robots operating in dense crop rows or greenhouse environments, where excessive control oscillation can impair steering precision and compromise task performance [80,81,82].

Another line of improvement integrates disturbance observers (DOB) into SMC frameworks to enhance robustness under uncertain operating conditions. By estimating external disturbances and internal variations—such as lateral slip, actuator degradation, or steering resistance—DOB-based SMC can provide more accurate and stable control in real-time scenarios [83,84]. In particular, high-fidelity DOBs have demonstrated effective compensation for slippage and force fluctuations in tracked UAGVs navigating gravel, muddy, or sloped terrains [85]. Nevertheless, practical implementation remains challenged by sensitivity to sensor noise and the need for precise bandwidth tuning.

To achieve faster and more predictable convergence, terminal and fixed-time sliding mode control (FTSMC) have been proposed. These approaches guarantee finite-time or fixed-time convergence regardless of the initial error magnitude, which is particularly valuable for rapid stabilization tasks such as headland turning or obstacle avoidance [86,87]. Recent designs combine nonsingular terminal sliding mode controllers (TSMCs) with disturbance estimators and integral sliding surfaces, achieving centimeter-level tracking accuracy in field experiments while maintaining robustness during abrupt terrain transitions [88,89]. Nonetheless, the requirement for higher-order derivative estimations and Lyapunov-based designs increases computational demands.

Finally, composite control strategies that integrate SMC with other methodologies—such as MPC, backstepping, or prescribed performance control—have gained traction. These hybrid controllers expand control horizons, embed future path information, and enforce error evolution within predefined performance bounds [90,91,92]. Such systems have demonstrated high precision and resilience in large-scale agricultural applications involving dynamic payloads, terrain gradients, and high-curvature trajectories. However, they typically depend on high-fidelity modeling and require significant onboard computation, which may limit their use in lightweight or resource-constrained platforms.

Table 6. Comparison of representative SMC-based strategies for agricultural path tracking.

SMC Method	Key Advantages	Main Limitations	References
CSMC	High robustness to matched disturbances; low computational cost	Chattering; limited adaptability to dynamic slip or mismatched disturbance	[71,74]
ASMC	Improved disturbance rejection and adaptability in uncertain	Difficult to cope with fast-changing conditions	[75,76,77]
Fuzzy-SMC	Smoother control action; reduced chattering	Design complexity; fuzzy rule tuning requires domain knowledge	[78,79,80,81,82]
DOB-SMC	Improved performance under unknown/mismatched disturbances; lower gain demand	Sensitive to noise and observer parameters	[83,84,85]
TSMC/FTSMC	Fast response; performance guaranteed within finite bounds	Relies on precise gain design; increase control effort	[86,87,88,89]
Composite SMC	High flexibility; capable of handling multi-source uncertainty and constraint conditions	Design complexity; increase control effort	[90,91,92]

summarizes representative SMC-based path-tracking strategies applied in UAGVs.

4.5. Model Predictive Control

MPC has emerged as a promising control strategy in UAGV applications, particularly due to its ability to anticipate future system behavior and handle multi-constraint optimization in real time [93,94]. These features are especially valuable in complex agricultural scenarios where vehicles must operate on uneven terrain, variable soil conditions, and under intermittent traction disturbances. The fundamental objective of MPC is to minimize the deviation between the predicted system trajectory and a predefined reference path, while ensuring compliance with dynamic constraints on system states and control inputs. Typically, this objective is formalized through a quadratic cost function, expressed as

$$\underset{u}{min}\sum _{k=0}^N\left(x_k^{T}Q{x}_k+u_k^{T}R{u}_k\right)$$

(14)

where x_k and u_k denote the predicted state and control input at time step k, while Q and R are positive-definite weighting matrices that balance tracking accuracy and control effort (Figure 13).

4.5.1. Linear MPC

Linear model predictive control (LMPC) is widely applied in UAGVs due to its computational efficiency and ease of implementation. By utilizing linear or linearized system models, LMPC enables real-time trajectory tracking under structured or moderately varying field conditions [95,96].

Several studies have adopted linear time-varying (LTV) MPC to enhance adaptability in complex agricultural environments. To handle significant model parameter variation during autonomous navigation tasks, an LTV-MPC controller was introduced in [97] for unmanned tractors. This method demonstrated strong robustness, tolerating over 10% parameter deviation while maintaining an average lateral error of 7 cm at a forward speed of 2.5 m/s. In response to the challenges posed by external disturbances and parameter uncertainties, ref. [98] proposed a unified augmented linear MPC framework that integrates both kinematic and dynamic vehicle models, resulting in substantially improved control robustness. To further enhance performance under low-adhesion conditions and complex terrain, a hierarchical control architecture was developed in [99], combining an upper-layer LTV-MPC for optimal path tracking with a lower-layer fuzzy PID controller for slip mitigation.

4.5.2. Nonlinear MPC

Nonlinear model predictive control (NMPC) has been widely applied in UAGVs for path tracking due to its ability to manage system nonlinearities, multivariable interactions, and real-time constraints [100]. Compared to linear MPC, NMPC offers improved performance in handling actuator dynamics, nonholonomic constraints, and environmental disturbances. Its effectiveness has been demonstrated across various platforms. For obstacle-dense environments, unified NMPC frameworks integrate tracking and avoidance, achieving cross-track errors under 5 cm [101]. Other formulations provide direct control of specific body points, supporting tasks like swath alignment and precision spraying [102].

To improve robustness, adaptive feedback and disturbance rejection mechanisms have been integrated, yielding over 30% accuracy improvements compared to LMPC under perturbations [103]. In structured row-following tasks, NMPC has been applied to novel platforms like gantry tractors [104] and skid-steered robots [105], where real-time trajectory planning must account for both motion constraints and obstacle avoidance. Computational efficiency is maintained through strategies like real-time iteration (RTI) and potential field integration.

4.5.3. Adaptive MPC

Adaptive model predictive control (AMPC) enhances conventional MPC by dynamically adjusting model parameters, prediction horizons, or weighting matrices based on real-time vehicle states and environmental conditions [106,107]. This flexibility is particularly beneficial in agriculture, where terrain, curvature, and speed vary frequently. One common approach is adaptive time-domain tuning, where prediction and control horizons are modified according to operating speed or path geometry. Similar improvements have been achieved in rear-wheel-steered systems using fuzzy logic and particle swarm optimization (PSO) to tune key MPC parameters [108].

Another strategy focuses on adaptive model updating. Techniques like adaptive forgetting factor recursive least squares (AFFRLS) continuously adjust tire stiffness and mass estimates based on road adhesion feedback, reducing lateral error by 67% [109]. To reduce computation without sacrificing accuracy, event-triggered AMPC has been proposed. In autonomous mowing tasks, curvature-based preview estimation combined with fuzzy logic adjusts prediction horizons in real time, maintaining heading errors within 0.13 rad and computation times under 5 ms [110].

4.5.4. Robust MPC

Robust model predictive control (RMPC) has been widely applied in agricultural environments to address trajectory tracking under terrain variability, wheel slip, and modeling uncertainties [111]. By explicitly accounting for bounded disturbances and unmodeled dynamics, RMPC enhances system stability compared to conventional MPC. In complex terrains such as slopes and orchards, hybrid RMPC approaches have been proposed, incorporating fuzzy anti-slip control [112] or slope compensation based on real-time force estimation [113], achieving tracking errors below 5 cm.

To manage model uncertainties more rigorously, tube-based RMPC constructs a bounded control region around the nominal trajectory, ensuring robust constraint satisfaction. This has been successfully applied in tractor–trailer coordination and off-road robotics [114]. Advanced designs integrate disturbance observers or sliding mode control to further enhance disturbance rejection [115], while min–max formulations improve tracking under uncertain slip conditions [116].

4.5.5. Summary

In summary, MPC has shown great promise in UAGV navigation, with each variant offering distinct advantages for agricultural field operations. Linear and linear time-varying MPCs remain attractive for structured environments due to their efficiency, but their performance degrades under slip and nonlinear terrain dynamics. NMPC addresses these limitations through the direct modeling of multivariable and actuator constraints, enabling better performance in curved paths, sloped terrains, and obstacle-rich environments–though often at the cost of higher computational demand. AMPC improves generalization by tuning horizons or updating models in response to real-time changes in soil, speed, or curvature. It is particularly suited to fields with variable traction or payload shifts. RMPC further strengthens system reliability under bounded uncertainties, making it effective for operations on muddy, sloped, or low-adhesion farmland.

Table 7. Comparison of representative MPC-based strategies for agricultural path tracking.

MPC Variant	Key Advantages	Main Limitations	References
LMPC	Handles constraints and multivariable couplings; low computational burden; suitable for real-time use	Limited handling of nonlinearity and strong disturbances	[97,98,99]
NMPC	Handles constraints and multivariable couplings; higher precision	High computational load; sensitive to model mismatch and tuning	[100,101,102,103,104,105]
AMPC	Improved adaptability to curvature, speed, terrain; balances tracking and stability; supports real-time tuning	Parameter tuning complexity; depends on reliable state estimation	[106,107,108,109,110]
RMPC	Strong resilience to slip, delay, and model mismatch; ensures constraint satisfaction under uncertainty	Increased design complexity; may be overly conservative; high computational cost in tube-based settings	[111,112,113,114,115,116]

provides a comparative summary of the reviewed MPC strategies–LMPC, NMPC, AMPC, and RMPC. Each method presents distinct advantages tailored to specific operational requirements, from computational efficiency to resilience under uncertainty.

4.6. Learning-Based Path-Tracking Control Methods

In recent developments, learning-based control has been explored as an alternative or supplement to traditional model-based methods in agricultural navigation. These approaches are particularly suited to handling nonlinearities, terrain variability, and disturbances that are difficult to model explicitly [117,118]. Techniques such as reinforcement learning, supervised learning, and hybrid frameworks combining data-driven models with conventional control have shown potential in improving adaptability and decision-making under uncertain conditions.

4.6.1. Reinforcement Learning Based Control

Reinforcement learning (RL) has shown promise in deriving control policies through interaction with the environment, without relying on explicit vehicle dynamics models. To address computational constraints, a dynamic self-triggered RL framework was introduced, maintaining robust tracking under disturbances while minimizing resource usage, and validated via second-order sliding mode control and Lyapunov analysis [119]. Additionally, adaptive dynamic programming (ADP) with critic neural networks has been used to iteratively solve the Hamilton–Jacobi–Isaacs equation, enabling control under high uncertainty [120].

4.6.2. Policy Optimization and Hybrid Architectures

Beyond value-based methods, policy-based reinforcement learning–such as proximal policy optimization (PPO)–has been applied to tractor–trailer systems by incorporating B-spline lane representations and tailored reward functions to address articulated dynamics [121]. Supervised RL has also been explored, where a linear parameter-varying (LPV) controller oversees the RL agent’s actions, providing a fallback under sensor noise or instability [122]. These approaches reflect a shift from manual rule design to adaptive, experience-driven strategies suited for the uncertain and variable conditions of agricultural environments.

5. Integration of Proximal Sensors with UAGVs

Proximal sensors mounted directly on UAGVs provide the immediate perception required for safe and accurate navigation. Unlike orbital or airborne sensing, which supplies large-scale geospatial information, proximal sensing offers real-time, high-resolution data at the field scale. These sensors constitute the core input for path planning, obstacle avoidance, and trajectory tracking in unstructured agricultural environments (Figure 14).

Vision-based sensors were the earliest to be adopted on agricultural vehicles, owing to their low cost and ability to capture rich visual information. Early work applied monocular and stereo vision to row detection and guidance [125]. Zhang et al. proposed a binocular-vision-based system combining Census-transform crop row detection, pure pursuit path planning, and PID steering control, achieving high precision under cotton field conditions [123]. More recently, deep learning approaches such as YOLOv8 have been applied to vineyard row detection under GNSS-denied conditions [124]. These studies show that vision remains a versatile and foundational tool for UAGV navigation. Figure 15 illustrates representative deployments of LiDAR sensors in UAGVs, supporting tasks such as obstacle detection, canopy profiling, and localization under variable field conditions.

LiDAR has more recently gained prominence for UAGVs due to its robustness under variable illumination and canopy conditions. Unlike cameras, LiDAR provides direct geometric measurements of the environment, enabling reliable row detection, obstacle avoidance, and trajectory planning. Jiang et al. [126] presented a 2D LiDAR-based trunk detection system for orchard spraying robots, while Xia et al. [127] integrated 3D LiDAR-SLAM with point cloud positioning for navigation in trellis orchards. Multi-sensor fusion frameworks often leverage LiDAR as the structural backbone of perception, highlighting its growing role in precise path-tracking control.

Overall, proximal sensing technologies provide the bridge between perception and control in UAGVs. While cameras laid the foundation for visual navigation, LiDAR has strengthened robustness and accuracy under challenging field conditions. Future systems are expected to integrate both modalities, combining the affordability and spectral richness of vision with the geometric precision of LiDAR to achieve reliable autonomy in diverse agricultural scenarios.

6. Applications of Control Strategies in Different Agricultural Scenarios

In practical agricultural operations, the effectiveness of tracking control strategies depends not only on their theoretical performance but also on their adaptability to the environmental characteristics and operational constraints of specific application scenarios. UAGVs are often deployed under diverse and complex working conditions, which vary considerably in terms of terrain structure, crop morphology, and task complexity. These variations directly affect the performance and stability of the control systems. Among the numerous agricultural scenarios, four working environments are particularly representative: drylands, paddy fields, orchards, and greenhouses. These scenarios encompass a wide range of typical agricultural conditions and have been extensively studied in both experimental research and commercial applications. This section reviews the implementation and practical application of control strategies in these representative environments, with the aim of providing insights into the applicability, robustness, and limitations of various control approaches under real-world agricultural conditions.

6.1. Deployment of Control Strategies in Dryland Fields

Dryland fields are typically used for cultivating crops such as wheat, maize, and soybeans. These environments are characterized by firm terrain and relatively regular path structures. However, large-scale machinery operations still encounter control challenges due to long-range path drift, limited headland turning space, and variable traction conditions. These factors necessitate control strategies that are not only precise but also capable of adapting to dynamic field conditions (Figure 16).

To address these challenges, a variety of control methods have been applied in dryland scenarios. A feedforward PID controller combined with a look-ahead Ackermann model was implemented on a rear-wheel-steered combine harvester, achieving an average lateral deviation of 5 cm on straight paths, and improving headland maneuverability through a three-cut turning scheme [128]. To overcome the limitations of the PP algorithm on uneven terrain, an optimal goal point selection method was introduced, enabling adaptive look-ahead distance adjustment and reducing tracking errors by more than 20% [129]. For whole-field navigation tasks, a fuzzy logic-enhanced Stanley controller was proposed, which dynamically tuned the control gain and improved both straight-line and turning accuracy compared to the standard model [130]. An improved pure pursuit algorithm using an enhanced sparrow search optimization was developed to dynamically tune speed and look-ahead distance, reducing lateral error by more than 50% compared to the standard model [131].

In summary, classical geometric controllers remain effective for structured and low-speed dryland operations. However, their performance tends to degrade in the presence of terrain variability or strict maneuvering constraints. Hybrid strategies such as fuzzy logic and model predictive control offer enhanced adaptability and robustness, and future developments should focus on integrating terrain-aware adaptation mechanisms and long-range error compensation for improved performance in large-scale dryland navigation.

6.2. Deployment of Control Strategies in Paddy Fields

Compared to dryland environments, paddy fields pose greater challenges to autonomous agricultural machinery due to waterlogged terrain, low surface adhesion, and variable sinking depths. These factors often result in wheel slip, lateral deviation, and unstable motion, particularly during complex maneuvers. Therefore, control strategies must be both robust to disturbance and capable of maintaining acceptable tracking accuracy under such dynamic field conditions (Figure 17).

Several studies have proposed control methods tailored to these challenges. A compound fuzzy PID controller was applied to a rear-steered rice transplanter, achieving an average lateral deviation in both straight and turning conditions of under 5 cm, demonstrating effectiveness in a moderately disturbed environment [132]. To enhance robustness under stronger nonlinearity, an adaptive sliding mode controller with RBF neural networks was introduced, significantly reducing the lateral and heading deviations compared to classical approaches [133]. In addition, MPC incorporating pose correction was used to address misalignment between machine and implement, achieving an average absolute error of 3.3 cm and effectively suppressing fluctuation due to attitude changes [134]. For highly unstructured or partially submerged paddy fields, a sliding mode controller with an extended disturbance observer enabled real-time compensation of slip effects and improved tracking robustness on 4WSS platforms [135].

In summary, fuzzy-PID and pose-corrected MPC methods are suitable for structured paddy environments with moderate slip, while adaptive and observer-enhanced sliding mode control demonstrates superior performance under high uncertainty. Future developments may benefit from integrating terrain perception with adaptive control to further enhance field robustness and real-time tracking stability.

6.3. Deployment of Control Strategies in Orchard Environments

Orchard environments present complex operational conditions for autonomous agricultural vehicles due to spatial constraints, frequent GNSS signal loss, uneven terrain, and non-standardized row geometry. These challenges often induce cumulative lateral drift, degraded localization accuracy, and delayed actuation responses—particularly in low-speed path-following scenarios involving towed implements such as sprayers. As a result, control systems must exhibit high adaptability to environmental variability and maintain trajectory precision under partial observability and nonlinear terrain disturbances (Figure 18).

To accommodate these demands, multiple control frameworks have been proposed. A navigation controller based on Double Deep Q-Network (Double DQN) was developed using a virtual radar model to encode spatial path-relative features. This learning-based approach demonstrated robust performance in orchard navigation, with field trials showing that lateral deviations were constrained within 0.27 m and reduced by more than 50% compared to conventional methods [136]. Complementarily, an ADP method employing a critic neural network was introduced to compute near-optimal control policies by solving the Hamilton–Jacobi–Isaacs equation online. This structure allowed the system to compensate for unmodeled dynamics such as wheel slippage and heading drift [120]. Moreover, for articulated vehicles commonly deployed in orchard settings, an adaptive model predictive control scheme integrating genetic optimization was proposed. It dynamically tuned the control horizon based on velocity and curvature, achieving up to 67.8% reduction in maximum lateral deviation across various path types [137].

In conclusion, double DQN offers an effective solution for constrained, perception-limited orchard environments, while ADP enhances disturbance rejection through online policy refinement. For articulated configurations and complex geometries, adaptive MPC yields superior path adherence by accommodating trajectory curvature and steering dynamics. Future work may focus on incorporating semantic terrain understanding and multi-modal fusion to further enhance decision making under orchard-specific uncertainties.

6.4. Deployment of Control Strategies in Greenhouse Agriculture

Greenhouse environments are relatively structured compared to open-field agricultural settings, often characterized by flat ground and predefined aisle layouts. Nevertheless, due to narrow corridors, small turning radii, and the high precision required by densely planted crop rows, they pose significant challenges for path-tracking control. These constraints demand accurate and responsive control strategies capable of maintaining stable trajectory tracking within tight spatial margins (Figure 19).

Various control strategies specifically designed for greenhouse applications have been proposed to cope with spatial constraints and precision requirements. A wireless positioning–based navigation and path tracking method for greenhouse mobile platforms achieved high positioning accuracy and reliable tracking performance [138]. In transplanting tasks involving ridge alignment, an ultrasonic ridge-tracking framework was developed by combining signal filtering and fuzzy PP control. The proposed method effectively suppressed measurement noise and achieved lower lateral deviations than conventional pursuit models, validating its suitability for structured greenhouse layouts [139]. Furthermore, a variable look-ahead PP strategy was implemented on a 4WIS–4WID platform, where the look-ahead distance was adaptively adjusted based on real-time tracking errors. Field experiments confirmed enhanced stability and reduced steady-state error compared to fixed-distance methods, highlighting its potential for improving trajectory convergence in greenhouse navigation [140].

From a control perspective, the design of path-tracking controllers for greenhouse applications must account for several unique constraints. The limited aisle width and frequent turning operations necessitate high-precision trajectory regulation, especially in the presence of tight boundary conditions. Controllers must strike a balance between responsiveness and smoothness to avoid abrupt steering or oscillations in confined environments. Moreover, due to the relatively predictable and repetitive nature of greenhouse layouts, adaptive mechanisms—such as fuzzy logic or variable parameter tuning—can be effectively leveraged to enhance control flexibility without incurring high computational costs. Overall, robust tracking under tight curvature constraints, minimal steady-state error, and real-time adaptability to slight geometric variations are critical requirements for ensuring reliable autonomous navigation in greenhouse settings.

7. Review Results

Based on the systematic literature analysis, the key findings of this review can be synthesized across modeling strategies, control algorithms, and agricultural application scenarios. Figure 20 presents a radar chart that benchmarks representative path-tracking methods across six evaluation dimensions, while

Table 8. Comparison of representative path-tracking control methods for UAGVs.

Method	Applicable Model	Advantages	Limitations	Applicable Scenarios	Comments
PID	Kinematic/dynamic model	Simple design; easy to implement; low computational demand	Sensitive to tuning; poor performance under slip, strong disturbances, or nonlinear dynamics	Straight row-following in greenhouses or flat dryland fields	Widely used baseline; relevant for low-cost platforms
PP	Geometric model	Geometric simplicity; intuitive path following; robust at low speeds	Accuracy depends on lookahead distance; limited in sharp turns; not robust to slip	Orchard row guidance; greenhouse navigation; moderate-structured fields	Well-studied; often benchmarked against other geometric controllers
Stanley	Geometric model	Effective for lane/row following; robust at moderate speeds; stable in structured rows	Oscillations in high-curvature or noisy GPS; limited robustness under extreme slip	Row-crop navigation with GNSS; vineyard/field spraying	More common in autonomous driving; fewer agricultural-specific adaptations
SMC	Kinematic/dynamic model	Strong robustness to disturbances and model mismatch; nonlinear handling	Chattering phenomenon; potential actuator wear; requires careful switching design	Sloped terrain; paddy fields with slip and soil–vehicle interaction	Variants with boundary-layer or higher-order SMC can reduce chattering
MPC	Kinematic/dynamic model	Explicit constraint handling; predictive optimization; high tracking accuracy	High computational demand; sensitive to model accuracy; requires solver	Variable-rate spraying; headland turning; integrated planning-control tasks	Active research trend; field-ready with fast solvers and embedded hardware
Learning-based	Flexible: can be combined with geometric, kinematic, or dynamic models	Adaptive to complex, unstructured environments; can leverage vision/ML datasets	Requires large, diverse datasets; risk of overfitting; limited explainability	GNSS-denied orchards; under-canopy navigation; crop-row detection via vision	Promising but still experimental; often combined with classical controllers

expands the comparative framework to include their applicable modeling assumptions, advantages, limitations, and deployment in typical agricultural settings. The results indicate that classical controllers such as PID, PP, and Stanley remain popular due to their simplicity, efficiency, and extensive validation in structured environments (e.g., greenhouses and flat dryland fields), but they provide limited robustness under disturbances or slip. Advanced strategies such as SMC and MPC demonstrate stronger robustness and constraint-handling capabilities, making them more suitable for challenging conditions such as paddy fields or sloped terrains, although issues such as chattering and high computational demand restrict their field adoption. Learning-based methods exhibit promising adaptability for unstructured environments such as orchards or GNSS-denied conditions, yet they remain highly dependent on large, diverse datasets, and face challenges in terms of generalization and practical maturity.

8. Conclusions and Future Directions

This review has provided a comprehensive analysis of path-tracking control methods for UAGVs, focusing on modeling strategies, control algorithms, and system performance. Geometric, kinematic, and dynamic models each offer distinct trade-offs between simplicity and fidelity. However, their effectiveness is often constrained by the unstructured and highly variable conditions typical of agricultural environments. While geometric and kinematic models remain popular due to their low implementation cost, they typically require site-specific calibration and struggle with non-ideal effects such as slip– and soil–vehicle interaction. In contrast, dynamic and extended models offer greater accuracy and robustness but at the expense of computational complexity. Integrating physics-based modeling with data-driven enhancements—such as disturbance observers and terrain-aware parameter estimation—holds promise for achieving both accuracy and adaptability in real-world deployments. Conventional approaches such as PID, PP, and Stanley methods remain prevalent due to their simplicity and low computational overhead. However, their performance degrades under significant disturbances or nonlinear dynamics. More advanced strategies such as SMC and MPC have demonstrated improved robustness and precision, especially in challenging scenarios involving slope, slip, and dynamic loading. Yet, the real-time deployment of these methods is often hindered by high computational demands. To address this, recent research has explored adaptive and learning-based frameworks—including reinforcement learning and hybrid MPC schemes—that offer enhanced flexibility and resilience. Advances in model compression, edge computing, and real-time optimization will be critical to translating these high-performance methods into practical, embedded agricultural systems.

Looking forward, several research directions deserve deeper exploration from the perspectives of modeling and control. To provide a clear and structured outlook, Figure 21 summarizes the key challenges identified in this review and maps them to promising research directions, serving as a visual roadmap for future work.

• High-fidelity modeling under multicondition variability: Most existing path-tracking controllers for UAGVs are built on simplified two-degree-of-freedom kinematic models. These abstractions are often insufficient to capture the complex dynamics of agricultural vehicles, particularly under variable load, traction loss, or aggressive maneuvers. Compared to autonomous road vehicles, UAGVs operate in harsher, less predictable environments. Future research should focus on modeling the coupled longitudinal–lateral–vertical dynamics, incorporating slip–soil interaction, and real-time parameter adaptation. Techniques such as online dynamics identification, hybrid physics-informed learning, and modular modeling can bridge the gap between model fidelity and control feasibility.
• Learning-based hybrid control methods: Most existing control strategies have been validated primarily in simulations or under idealized conditions. In contrast, real agricultural deployments—particularly when operating on slopes, wet grasslands, or muddy paddy fields—are characterized by pronounced uncertainty, strong nonlinearities, and diverse external disturbances. To remain effective in such adverse environments, controllers must be capable of maintaining stability and performance despite rapidly changing operating conditions. A promising future direction is the development of learning-based hybrid control architectures, for example, by combining learning with MPC or SMC, which can integrate the adaptability of data-driven approaches with the robustness and constraint-handling capabilities of classical control frameworks, thereby enhancing both robustness and generalization across complex agricultural scenarios.
• Integrated planning and control under physical constraints: Improving controller performance alone is insufficient when the upstream path planner fails to generate feasible or dynamically trackable trajectories. The mismatch between planned paths and UAGV actuation constraints can lead to poor stability, large tracking errors, or infeasible maneuvers. To address this, tighter coupling between local path planning and trajectory control is needed. MPC-based frameworks provide a natural solution by unifying prediction, multi-objective optimization, and constraint handling. Future research should explore integrated motion planning and control architectures that account for terrain perception, vehicle dynamics, and task-specific constraints (e.g., weeding, spraying) in real time.
• Integration of Multiple sensing with UAGVs: While UAGVs currently rely mainly on proximal sensors and GNSS/RTK, orbital sensing can provide valuable large-scale geospatial information. Satellite imagery and radar data enable the delineation of field boundaries, identification of crop variability, and detection of soil and vegetation conditions across entire fields. By integrating such global information with local perception, UAGVs could adapt trajectories, treatment intensity, and task scheduling according to spatial prescriptions. This multi-layered approach would support more informed path planning and coordinated operation across multiple machines. As satellite temporal and spatial resolutions continue to improve, orbital sensing is expected to play a growing role in seasonal management and the long-term optimization of UAGV operations.
• Agronomic-aware control strategies: Control design for UAGVs should be aligned with agronomic goals. Precision operations require that control strategies account for spatial variability in crops, soil heterogeneity, and task-specific constraints such as row spacing, planting density, or treatment zones. Future controllers should integrate agronomic knowledge into decision making—either through adaptive parameterization or modular task-oriented control layers. Moreover, as agricultural norms evolve with climate and regional factors, control strategies must also be flexible to adapt across seasons, crop types, and field geometries, thereby promoting machine–agronomy synergy.