Application of Navigation Path Planning and Trajectory Tracking Control Methods for Agricultural Robots

Abstract

Autonomous navigation is a core enabler of smart agriculture, where path planning and trajectory tracking control play essential roles in achieving efficient and precise operations. Path planning determines operational efficiency and coverage completeness, while trajectory tracking directly affects task accuracy and system robustness. This paper presents a systematic review of agricultural robot navigation research published between 2020 and 2025, based on literature retrieved from major databases including Web of Science and EI Compendex (ultimately including 95 papers). Research advances in global planning (coverage and point-to-point), local planning (obstacle avoidance and replanning), multi-robot cooperative planning, and classical, advanced, and learning-based trajectory tracking control methods are comprehensively summarized. Particular attention is given to their application and limitations in typical agricultural scenarios such as open-fields, orchards, greenhouses, and hilly slopes. Despite notable progress, key challenges remain, including limited algorithm comparability, weak cross-scenario generalization, and insufficient long-term validation. To address these issues, a scenario-driven “scenario–constraint–performance” adaptive framework is proposed to systematically align navigation methods with environmental and operational conditions, providing practical guidance for developing scalable and engineering-ready agricultural robot navigation systems.

1. Introduction

Autonomous navigation is a core enabler of smart agriculture, supporting unmanned operations such as spraying, seeding, harvesting, and intra-field logistics [1,2,3]. Within the navigation pipeline, path planning provides collision-free and efficiency-oriented motion references, while trajectory tracking control executes these references under nonlinear vehicle dynamics and field disturbances [4,5].

Agricultural environments are highly heterogeneous (e.g., open fields, orchards, greenhouses, and sloped terrain) and introduce practical constraints such as occlusions, GNSS-denied segments, soil-induced slippage, narrow passages, and dynamic obstacles, which jointly challenge both planning and control robustness [6,7,8]. Figure 1 provides a reference navigation stack for agricultural robots, organized into sensing, decision, and control layers and linked to task execution. The sensing layer integrates GNSS/RTK, IMU, cameras, and LiDAR to support localization, perception, and multi-sensor fusion. The decision layer covers global planning (coverage and point-to-point), local planning with real-time obstacle avoidance, hybrid planning, and multi-robot coordination for large-scale operations. The control layer translates planning outputs into actuator commands through classical controllers, optimization-based controllers, and learning-enhanced/adaptive controllers. This layered abstraction is used throughout the subsequent sections to structure the taxonomy and to clarify how planning and tracking methods interact under different agricultural scenarios [9,10,11,12].

1.1. Scope, Novelty, and Contributions

Recent surveys (2020–2025) have typically focused on (1) general mobile-robot path planning; (2) agricultural ground-robot planning; (3) vision-based navigation; (4) modeling and control for path tracking. However, planning and control are often discussed in isolation, and the linkage between algorithmic characteristics and scenario-specific constraints is not always made explicit for agricultural deployment [13,14]. To address this gap, this review contributes the following:

(1)

A coupled “decision–execution” synthesis that jointly reviews planning and tracking control methods as an integrated navigation stack;

(2)

A structured comparison matrix that summarizes strengths, limitations, and deployment constraints of major method families;

(3)

A scenario–constraint–performance adaptability framework that maps algorithm selection to four representative agricultural scenarios (open-fields, orchards, greenhouses, and hilly slopes);

(4)

A consolidated discussion of evaluation metrics, reproducibility barriers, and engineering maturity requirements for long-term field deployment.

Compared with recent review studies published, which typically focus on either path planning algorithms, perception-driven navigation, or control strategies in isolation, this review explicitly emphasizes the interaction between agricultural scenarios and navigation method selection. Existing surveys generally classify algorithms by technical categories, but seldom analyze how environmental constraints (such as field geometry, crop density, terrain slope, and GNSS availability) systematically affect the suitability, limitations, and performance boundaries of planning and control methods.

The novelty of this review therefore lies not in proposing new algorithms, but in providing a scenario-oriented synthesis that connects algorithmic characteristics with practical deployment conditions. By introducing a “scenario–constraint–performance” analytical framework, this work clarifies why certain methods that perform well in open-field environments may become unsuitable in orchards or greenhouses, and why advanced controllers do not necessarily translate into better field performance under resource-constrained platforms. This perspective is intended to support informed method selection and system-level design decisions for agricultural robot navigation, rather than algorithmic benchmarking alone.

1.2. Review Methodology

To enhance transparency and reproducibility, this review followed a PRISMA-style literature identification and screening workflow. We searched two databases (Web of Science and EI Compendex) for studies published between 2020 and 2025, using keyword combinations centered on “path planning” AND/OR “trajectory tracking”, optionally combined with “agriculture”, “agricultural robot”, “field robot”, “navigation”, and “agricultural machinery”. The overall thematic scope of the retrieved literature is illustrated by the keyword co-occurrence network in Figure 2.

(1)

Information sources and record accounting. The initial search retrieved 740 records in total, including 654 records from Web of Science and 86 records from EI Compendex. EI Compendex was used to supplement EI-indexed Chinese literature not covered by Web of Science. Because the Web of Science search involved iterative refinement of keyword combinations, partial overlaps occurred across different query rounds. All records were therefore exported into a single library and deduplicated using title and DOI matching within Web of Science. After removing 124 duplicate records, 616 unique records (N₁ = 616) remained for further screening;

(2)

Eligibility criteria. Studies were included if they (i) proposed, evaluated, or implemented path planning and/or trajectory tracking control methods applicable to agricultural robots or agricultural machinery navigation, and (ii) provided sufficient technical details, including algorithmic descriptions, evaluation settings, and outcome metrics, to enable method-level comparison. Studies were excluded if they (i) were unrelated to mobile robotics or navigation, (ii) lacked technical interpretability due to insufficient methodological detail, or (iii) had no accessible full text for verification;

(3)

Screening procedure. Title and abstract screening excluded 471 records, primarily due to topic mismatch (e.g., non-agricultural robotics, UAV-only navigation, or purely theoretical planning without navigation implementation) or insufficient technical content. Full-text assessment was subsequently conducted on 145 articles (N₂ = 145). Among these, 50 studies were excluded because of ineligible research scope, missing methodological details, or unverifiable experimental results. The final corpus therefore comprised 95 included studies (N₃ = 95), including 80 papers retrieved from Web of Science and 15 EI-indexed Chinese studies (see Figure 3);

(4)

Data extraction and coding. For each included study, we systematically extracted information on scenario type (open-fields, orchards, greenhouses, and hilly slopes), sensing and localization modality, planning paradigm, controller type, robot platform, evaluation setting (simulation or field experiments), and key performance indicators. These attributes were used to qualitatively categorize and compare the included studies throughout the subsequent sections, enabling a structured analysis of methodological trends, representative solutions, and remaining research gaps.

To improve interpretability and reveal research concentration and gaps, the included studies were analyzed from two complementary perspectives. First, for methodological analysis throughout the review, the literature was categorized along four technical dimensions: (1) agricultural environment (e.g., open-fields, orchards, greenhouses, and hilly slopes); (2) robot platform (e.g., wheeled, tracked, articulated, or hybrid configurations); (3) sensor and localization modality (e.g., GNSS/RTK-based, vision-based, LiDAR-based, or multi-sensor fusion); and (4) dominant trajectory tracking or motion control strategy. Because many studies adopt hybrid sensing configurations and combined control strategies, individual works may be discussed under multiple categories where appropriate. This multi-dimensional classification is primarily used to organize the narrative technical analysis rather than to provide an exhaustive quantitative breakdown. Second, to provide a high-level overview of the research emphasis within the reviewed corpus, each study was additionally assigned to a primary research focus category according to its main contribution: path planning–focused, trajectory tracking/control–focused, or integrated planning-and-tracking studies. This simplified categorization, summarized in Figure 4, is intended to capture the dominant research objective of each work rather than its complete technical composition.

1.3. Paper Organization

The structure of this review is as follows: Section 2 provides a detailed overview of path planning algorithms, including full-coverage path planning, point-to-point navigation, local planning with real-time obstacle avoidance, hybrid planning, and multi-robot coordination strategies. Section 3 focuses on trajectory tracking control methods and modeling, systematically reviewing and comparing classical, advanced, and intelligent adaptive approaches. Section 4 analyzes the application and challenges of key navigation technologies in typical agricultural environments such as open fields, orchards, greenhouses, and sloped terrain. Section 5 outlines future development trends for navigation path planning and tracking control technologies in agricultural robots. Finally, Section 6 provides a summary of the entire paper.

2. Path Planning Algorithm

Path planning constitutes a key decision-making component in agricultural robot navigation, yet its practical effectiveness is strongly constrained by agronomic requirements, field geometry, and operational efficiency rather than purely geometric optimality [11]. Based on scope and purpose, path planning for agricultural robots is typically categorized into global and local path planning. Global path planning generates a route covering the entire operational area or connecting key points based on known global environmental information. Local path planning performs real-time adjustments and obstacle avoidance within the global path framework, utilizing real-time local environmental data (dynamic obstacles) [15]. Hybrid path planning strategies offer superior adaptability across diverse environments. Furthermore, multi-robot collaborative path planning has emerged as a key technology for enhancing efficiency in large-scale farmland operations.

2.1. Global Path Planning

Global path planning primarily falls into two categories: complete coverage path planning (CCPP), which aims to comprehensively cover the operational area; and point-to-point navigation planning, which seeks to achieve optimal movement between two designated points.

2.1.1. Complete Coverage Path Planning

Complete Coverage Path Planning (CCPP) stands as one of the most distinctive path planning methods in agricultural robotics. It generates a path that comprehensively covers the entire working area without repetition or omission, while minimizing non-working distance and time. Agricultural operations demand exceptional coverage integrity and efficiency—minor coverage gaps may lead to yield losses, whereas excessive overlap causes resource waste and soil compaction [16,17,18].

Early CCPP methods are primarily based on rule-driven geometric strategies, whose continued adoption in agricultural applications is largely motivated by engineering robustness and computational efficiency rather than theoretical optimality [19,20]. Patterns such as boustrophedon paths implicitly assume regular field boundaries and limited internal obstacles, under which they provide predictable motion behavior and low implementation cost. However, their effectiveness degrades rapidly when these assumptions are violated. Irregular boundaries, fragmented plots, or internal obstacles inevitably introduce redundant coverage and non-working travel, and such inefficiencies cannot be resolved through parameter tuning alone. Although field decomposition strategies partially alleviate this issue, their performance becomes highly sensitive to the quality of the decomposition itself. Consequently, geometric CCPP methods should be regarded as scenario-specific engineering baselines rather than universally optimal solutions, remaining most suitable for large-scale, structured farmland with stable geometric characteristics. Höffmann et al. [21] conducted in-depth research on CCPP methods for precision agriculture, proposing an efficient field decomposition and optimal path generation system that reduces soil compaction and enhances crop yields.

Coverage paths generated by the aforementioned geometric decomposition methods typically consist of two fundamental elements: parallel straight paths within the working area and headland turning paths. Parallel straight paths must be designed to strictly adhere to agronomic row spacing requirements to ensure complete coverage without overlap. Headland turning path planning is directly constrained by the agricultural machinery or robotic chassis structure (wheeled, tracked, Ackermann steering, skid steering) and its kinematic model. For instance, chassis employing Ackermann steering struggle to execute efficient fishtail turns, while robots with skid steering or four-wheel independent steering possess superior on-the-spot turning capabilities [22].

Optimizing turning strategies is key to enhancing the overall operational efficiency of CCPP. Common headland turning methods (see Figure 5) can be categorized into three types based on their geometric characteristics and maneuvering complexity: (1) U-turns, Ω-turns, and their smooth variants: These include standard U-turns and Ω-turns (or circular turns) with larger turning radii. Such turning paths are smooth and simple to execute but require the widest headland space; (2) Hook/Pear-Shaped Turns: This asymmetric composite maneuver involves a forward turn followed by a reverse turn to complete the headland reversal. It demands higher precision but requires less headland width than U-turns; (3) Fishtail Maneuver: Achieved through a series of small-angle forward and reverse swaying movements. Demands the highest maneuvering precision but requires the least headland width among all turning methods, making it particularly suitable for narrow headlands [23]. Therefore, selecting the optimal turning strategy involves a multi-objective trade-off between working width, minimum turning radius of the agricultural machinery, chassis structural characteristics, and dynamic performance. Advanced CCPP algorithms incorporate turning strategies as part of the optimization objective to minimize total non-working time and energy consumption.

To overcome the limitations of traditional geometric methods in complex environments, intelligent algorithms employ iterative search to find near-optimal coverage path sequences. Genetic algorithms (GA) simulate natural selection processes by encoding, crossing over, mutating, and selecting the visit order of path points, continuously evolving superior paths. This approach effectively handles irregularly shaped plots and reduces redundant paths. Simulated annealing (SA) mimics solid annealing processes, incorporating probabilistic jump mechanisms to avoid local optima, making it suitable for solving complex combinatorial optimization problems. Particle swarm optimization (PSO) [24] and ant colony optimization (ACO) simulate collective intelligence behaviors, finding global optima through information sharing and collaboration among particles/ants, particularly effective for large-scale, multi-constraint path planning problems. Liu et al. [25] applied an improved ACO to orchard mower path planning, significantly enhancing coverage while reducing time by 47.58%. However, the algorithm is sensitive to initial parameters and plot partitioning, requiring parallelization or partitioning strategies for online deployment to mitigate latency. The strength of these intelligent methods lies in their robust global search capabilities, enabling handling of highly complex field shapes and internal obstacles. But their high computational complexity makes real-time performance challenging, typically limiting their application to offline planning scenarios. CCPP algorithm selection requires balancing trade-offs based on specific application contexts. In regular large-scale field environments, geometric methods remain the preferred choice due to their simplicity and efficiency. For complex irregular plots, intelligent algorithms offer superior solutions.

2.1.2. Point-to-Point Path Planning

The goal of point-to-point navigation planning is to find the optimal (or suboptimal) collision-free path from a given starting point to a destination. The optimality criterion can be the shortest path length, minimum time consumption, or lowest energy expenditure. This type of planning is commonly used for agricultural robots moving between different fields.

Graph search algorithms form the foundation for solving point-to-point planning problems. Among them, Dijkstra’s algorithm stands as one of the most classic approaches, guaranteeing the discovery of the global shortest path within grid maps or graph structures. However, it requires traversing all nodes, resulting in low computational efficiency and unsuitability for large-scale environments. The A* algorithm builds upon Dijkstra’s by introducing heuristic functions to guide the search direction, significantly enhancing search efficiency. It has been successfully applied in navigating many regular agricultural plots. Nevertheless, A*’s performance heavily depends on heuristic function design, and in fields larger than 50 hectares, computational latency may exceed 30 s, making it difficult to meet real-time requirements. Furthermore, A* and Dijkstra are typically based on discrete grid maps, potentially generating paths that lack smoothness and fail to account for robotic kinematic constraints. The asterisk (*) suffix in these algorithms indicates the integration of heuristic functions to guide the search direction, fundamentally distinguishing them from blind search methods like Dijkstra.

Beyond static graph-search planners, incremental replanning algorithms such as D* and D* Lite address a core deficiency of A*-based methods in agricultural environments: the mismatch between static maps and continuously evolving field conditions. Rather than recomputing paths from scratch, D* Lite incrementally repairs only the locally affected regions of the cost map when discrepancies are detected by onboard perception [26]. This property is particularly valuable in agricultural settings, where environmental changes are frequent yet spatially localized, for example, temporarily parked machinery, fallen branches, or human workers entering the operation zone. In such cases, full replanning not only increases latency but may also induce unnecessary path oscillations. By contrast, incremental replanning preserves global path consistency while enabling fast local adaptation.

Nevertheless, the practical performance of D*-family algorithms is tightly coupled with map resolution, cost-update strategies, and sensor reliability. In muddy or vegetation-dense fields, noisy cost updates may trigger frequent replanning events, leading to unstable path switching unless hysteresis mechanisms or cost-smoothing strategies are introduced. Therefore, D* Lite should be regarded as a dynamic extension of A*, rather than a universally superior alternative, and is most effective when embedded within a perception-aware navigation stack.

Sampling-based algorithms construct paths by randomly sampling continuous state spaces, better addressing challenges in high-dimensional and continuous spaces. Rapidly Exploring Random Trees (RRT) and its optimized variant RRT* are representative examples [27]. RRT* is an asymptotically optimal algorithm that generates smooth paths satisfying robot dynamics constraints by continuously reconnecting to optimize path cost. Ye et al. [28] proposed an improved, kinematically constrained bidirectional RRT method, achieving maximum path length reductions of 13.94% and 12.20%, with planning time reductions of up to 61.41% and 68.84%. RRT* excels in search efficiency independent of environmental spatial dimensions, making it suitable for complex agricultural terrain. However, paths remain outcomes of random sampling, potentially suboptimal and featuring unnecessary meandering. Probabilistic Roadmap Method (PRM) offers a complementary sampling approach to RRT, characterized by a two-stage model: offline construction and online querying. First, it generates random path points covering free space and connects them into collision-free roadmaps. Subsequently, graph search algorithms rapidly solve paths between arbitrary start and end points within this fixed network. This method’s advantage lies in its ability to support multiple efficient queries after a single construction, making it particularly suitable for global planning in static environments with multiple tasks. However, it struggles to directly handle dynamic environments, and the computational overhead of online reconstruction is significant.

In agricultural environments with dense vegetation or repetitive row structures, the randomness of sampling-based planners may also introduce unnecessary path oscillations, increasing control burden during trajectory tracking.

Although RRT and PRM are categorized under global planning, their real-time capabilities also make them highly effective in local replanning. Subsequent Section 2.2 will not elaborate further on this aspect. When a global path is interrupted by a dynamic obstacle, a local RRT instance based on current sensor information can often be initiated to rapidly generate a local path that bypasses the obstacle, then reconnect to the global path. PRM focuses more on offline construction of a global roadmap, offering extremely high efficiency during online queries, but it lacks the flexibility of RRT.

2.2. Local Path Planning and Real-Time Obstacle Avoidance

Although global planning provides a nominally optimal route, agricultural field environments frequently violate its underlying assumptions due to dynamic obstacles, map inaccuracies, and unstructured terrain, making purely global solutions insufficient for safe operation. Local path planning handles more precise adjustments and real-time obstacle avoidance based on sensor data from the perception layer, while adhering to the overall direction of the global path [29]. This ensures the robot’s safe operation.

2.2.1. Perception and Multi-Sensor Fusion

The performance of local path planning and real-time obstacle avoidance heavily relies on the accuracy, completeness, and timeliness of environmental information provided by the perception system. Single sensors have inherent limitations and struggle to handle all scenarios. Multi-sensor fusion has become the standard configuration for agricultural robot perception systems [30,31,32]. By leveraging complementary information from different sensors, it enhances the overall robustness and reliability of the system under extreme operating conditions.

A representative sensing suite for agricultural robot navigation (see Figure 6) typically integrates complementary sensors to ensure robustness under variable illumination, dust, occlusion, and intermittent GNSS. Common configurations include (1) LiDAR, providing high-precision 3D point clouds for obstacle geometry and terrain structure; (2) stereo cameras, providing texture and semantic cues for crop-row detection and object classification; (3) IMU, providing high-rate inertial measurements for state propagation and motion stability; and (4) GNSS/RTK modules, providing absolute position and heading when signals are available. By fusing these heterogeneous streams, the system can generate reliable local cost maps and state estimates to support local planning and real-time obstacle avoidance.

LiDAR and stereo cameras undergo joint calibration to complement each other’s data, jointly constructing a high-precision local cost map around the robot and enabling semantic recognition of obstacles. As the core of state perception, the IMU provides motion prior for LiDAR and visual odometry, effectively correcting distortions and smoothing estimation results. GNSS/RTK provides an absolute pose reference when signals are available, correcting global errors caused by IMU drift and odometry accumulation. All perception and state information is ultimately unified within the fusion center, delivering reliable, continuous, and multi-level environmental and state representations for subsequent local planning algorithms. This constitutes the primary prerequisite for achieving safe and reliable autonomous navigation.

It should be noted that improvements in perception accuracy do not translate linearly into planning performance. Overconfident obstacle inflation caused by sensor noise may overly constrain some local planners, leading to conservative behavior and unnecessary detours, whereas underestimation of obstacle boundaries directly compromises safety. Consequently, perception–planning coupling, rather than sensor richness alone, becomes the key determinant of local navigation reliability in agricultural robots.

2.2.2. Classical Algorithm

Classical local planners remain attractive in agricultural robotics due to their low computational cost and ease of implementation on embedded platforms. However, their behavior is strongly shaped by heuristic scoring terms and parameter tuning, which can cause brittle performance when environmental structure changes (e.g., narrow orchard aisles versus open fields).

DWA (Dynamic Window Approach) explicitly accounts for dynamic constraints through velocity-space sampling, typically achieving good real-time responsiveness, but it may become conservative or oscillatory in cluttered, vegetation-dense scenes. APF (Artificial Potential Field) provides intuitive attraction–repulsion navigation but is prone to local minima and narrow-passage oscillations, which are common in orchards and greenhouses. VFH (Vector Field Histogram) method is comparatively robust to sensor noise and can yield smooth avoidance, yet their performance depends on reliable obstacle representation and may degrade when foliage produces fragmented returns. In practice, these methods are often deployed as local “reactive layers” combined with global guidance or hybrid planning to mitigate local-optimum failure modes.

2.2.3. Algorithm Based on Reinforcement Learning

Learning-based local planning like RL (Reinforcement learning) or DRL (Deep reinforcement learning) is increasingly explored for agricultural navigation because it can encode complex objectives (e.g., row-following constraints, obstacle avoidance, smoothness, and task efficiency) into a unified policy and can adapt to perception uncertainty and GNSS outages. Nevertheless, current evidence suggests three deployment barriers: (1) safety and reproducibility—training and evaluation protocols vary widely and negative/failure cases are underreported; (2) sim-to-real sensitivity—policies trained under limited soil/illumination/occlusion distributions may degrade substantially in the field; (3) compute and latency constraints—real-time inference is feasible, but training and tuning are expensive, and performance can be brittle under distribution shifts. Consequently, RL/DRL is most defensible in hybrid stacks where learning components improve local reactivity or perception-to-action mapping, while safety envelopes and fallback planners/controllers preserve stability and constraint satisfaction [33,34]. Fu & Yao [35] reported that RL obstacle avoidance strategies trained in simulation achieve collision-free navigation in over 90% of scenarios, though occasional strategy failures occur in limited real-world tests. Kang et al. [36] deployed a DRL local navigation strategy in narrow strawberry greenhouse aisles, achieving a 15–20% reduction in lateral trajectory error compared to traditional baselines. However, this approach exhibits strong dependence on training samples, and “Sim-to-Real” transfer requires domain randomization and online fine-tuning for stable operation in real environments.

In agricultural navigation, RL is modeled as a Markov Decision Process (MDP), defined by the tuple (S, A, P, R, γ). Key implementation steps include the following: (1) State Space (S) Construction: Fusing LiDAR range data (geometric constraints) with visually detected crop row headings (agronomic constraints) and robot kinematic states; (2) Reward Function (R) Shaping: Unlike general robotics, the reward function for agricultural robots must integrate a crop protection term: $R_{total}=w_1{R}_{goal}+w_2{R}_{smooth}-w_3{R}_{crop\_damage}$, where $R_{crop\_damage}$ imposes heavy penalties for any intrusion into the crop root zone, prioritizing operational safety over speed; (3) Training and Transfer: To bridge the “Sim-to-Real” gap caused by variable lighting and unstructured foliage, Domain Randomization is applied to texture and physics parameters in simulation. Furthermore, Curriculum Learning is adopted, allowing the agent to progress from simple empty-row following to complex scenarios with weeds and dynamic obstacles, ensuring policy convergence in high-dimensional environments.

The advantage of RL methods lies in their robust environmental adaptability, excelling at handling previously unknown or highly dynamic scenarios. However, RL training requires substantial data, and trained policies may fail in real environments due to state distribution discrepancies, posing safety risks. RL and DRL applications primarily focus on local planning. Yet through clever design of state spaces, action spaces, and reward functions, DRL can also learn long-range navigation strategies, enabling global path planning from start to finish. When incorporating global target location information into the state space, robots can learn to reason about key path points leading to the destination.

2.2.4. Algorithm Based on Fuzzy Logic

Fuzzy logic (FL) provides an effective method for handling imprecision and uncertainty in local obstacle avoidance for agricultural robots, making it suitable for field scenarios characterized by high sensor noise and ambiguous environmental information. Unlike traditional binary logic, FL permits continuous values between 0 and 1. A fuzzy logic obstacle avoidance system typically involves three steps: fuzzification, fuzzy reasoning, and defuzzification. First, precise inputs (such as distances to obstacles on the left and right sides, or lateral deviation from the global path) are converted into fuzzy linguistic variables (such as “near,” “medium,” “far”). Next, reasoning occurs based on a fuzzy rule base designed with expert knowledge (If the left obstacle is close and the right obstacle is far away, steer to the right.). Finally, the inferred fuzzy outputs are defuzzified into precise control quantities (such as specific steering angles or velocity commands).

FL does not rely on precise mathematical models, offering high robustness, computational efficiency, and ease of integrating expert knowledge. It is well-suited for embedded platforms with limited computational resources. Liu et al. [37] combined fuzzy control with an improved A* algorithm for track-based agricultural machinery guidance, demonstrating its capability to handle uncertainty. However, fuzzy logic is highly dependent on system performance; incomplete or poorly designed rule bases may lead to erratic behavior. While adaptive fuzzy systems or integration with learning algorithms can automatically adjust rules, this increases system complexity. FL is often combined with other methods (see Section 2.3) to form hybrid intelligent systems, enhancing overall performance.

It is particularly noted that intelligent methods such as RL and FL, described in Section 2.2.2 and Section 2.2.3, are not only applicable at the decision-making level for local path planning and obstacle avoidance. Their underlying principles are also widely applied in trajectory tracking control at the execution level, which will be discussed in detail in the next chapter (see Section 3.4). Their application focuses differ: in planning, they primarily address behavioral decisions under uncertainty, such as how to navigate around obstacles; in control, they emphasize parameter optimization and dynamic compensation, such as how to precisely execute steering commands.

2.3. Hybrid Path Planning Strategy

In agricultural environments, path planning faces highly complex, dynamic, and uncertain challenges. Single algorithms often struggle to balance multiple objectives such as global optimality, real-time obstacle avoidance, and computational efficiency. Hybrid path planning strategies can achieve more powerful and robust performance by effectively combining the strengths of different algorithms. Common hybrid approaches fall into three categories:

(1)

Hierarchical architectures integrating global and local planning. The global planner generates a coarse overall path based on prior maps, ensuring overall goal reachability and optimality; the local planner handles real-time perception data, managing dynamic obstacles and unknown terrain to achieve real-time obstacle avoidance and trajectory tracking;

(2)

Integration of optimization methods with search/sampling algorithms. As discussed in Section 2.1.2, these approaches leverage search algorithms for rapid exploration of high-dimensional spaces, then employ optimization algorithms to post-process initial paths. This achieves path smoothing, length reduction, and decreased motion cost;

(3)

Complementary integration of learning methods with traditional planning algorithms. Neural networks and reinforcement learning methods enhance traditional planning algorithms’ understanding and prediction capabilities in complex environments—such as identifying crop growth states, predicting dynamic obstacle intentions, or adapting to slippery/uneven terrain—while traditional planners maintain foundational stability and safety guarantees.

Beyond these three categories, FL is frequently integrated into hybrid systems to handle the pervasive uncertainty and subjective knowledge inherent in agricultural environments. Within local obstacle avoidance modules, FL can be combined with APF or DWA to reason about linguistic variables such as “distance-to-hazard severity” and “steering urgency,” generating smoother behavioral commands that align more closely with human experience. This approach performs particularly well when image recognition confidence is low or sensor noise is significant. Despite their superior flexibility, hybrid planning architectures introduce higher system complexity and integration cost, which may hinder deployment on resource-constrained agricultural platforms.

2.4. Multi-Robot Collaborative Path Planning

Multi-robot collaborative operations represent a key technology for enhancing the efficiency of large-scale farmland operations and reducing operational time. Multi-robot collaborative path planning involves assigning work areas to multiple robots and designing safe paths to minimize total operation time while preventing collisions between robots and ensuring operational completeness and balance. In recent years, numerous researchers have focused on multi-robot coordination: Lytridis et al. [38] demonstrated in a two-robot simulation that coordination could reduce operation time by 35% while maintaining individual robot accuracy. However, they noted that unstable real-world communication significantly impacts coordination benefits, necessitating the design of communication failure tolerance strategies. Wang et al. [39] conducted field tests with a four-robot formation, controlling collision probability to <1% via directional bounding box collision detection. Task completion time was reduced by 22% compared to greedy allocation, demonstrating engineering feasibility, though demanding high communication bandwidth and synchronization accuracy.

Existing methods are primarily categorized into three types based on decision-making structures (see

Table 1. Comparative Analysis of Multi-Robot Collaboration Methods.

Architecture	Core Mechanism	Optimality and Efficiency	Scalability	Robustness and Fault Tolerance	Application Scenarios
Centralized	Central server possesses global information and computes planning for all robots	Theoretically globally optimal, but computationally complex	Poor	Low	Small-scale, structured environments with robust communication infrastructure
Distributed	Robots make independent decisions based on local information and peer-to-peer negotiation	Typically locally optimal, with no guarantee of global performance and potential for oscillation	Excellent	High	Large-scale, dynamic environments with limited communication capabilities
Hybrid	Central node assigns high-level tasks, robots perform local autonomous planning	Balancing overall efficiency with local flexibility	Good	Medium	Medium-to-large-scale environments requiring both global coordination and localized responsiveness

):

(1)

Centralized Planning. A central control unit collects all environmental information and robot states, uniformly assigning tasks and planning paths. This approach theoretically achieves globally optimal solutions and avoids conflicts. However, it demands extremely high reliability and bandwidth from communication networks, places heavy computational burdens on the central unit, and carries single-point-of-failure risks. Failure of the central node or communication links can paralyze the entire system;

(2)

Distributed Planning. Without a central node, each robot autonomously makes decisions based on its own sensory data and limited communication with neighboring robots. Typical approaches include task allocation via auction mechanisms and collaborative decision-making using consensus algorithms. Distributed methods offer robust system resilience and scalability, remaining functional even with individual node failures. However, they struggle to guarantee global optimality and may even degrade system performance due to local decision conflicts. Huo et al. [40] developed a safety detection model and collaborative turning strategy for master-slave tracking operations, effectively mitigating collision risks during headland turns;

(3)

Hybrid Planning. Combining features of both approaches, a central node typically assigns macro-level tasks while robots autonomously plan local details. This scheme balances overall efficiency with local flexibility but requires more complex system design and clear delineation of decision-making responsibilities between central and local entities.

Notably, many reported performance gains are demonstrated under controlled communication conditions, whereas unstable wireless links in real farmland often limit the scalability of multi-robot coordination strategies. Method selection depends on specific scenarios: centralized approaches are suitable for small-scale static environments to pursue optimal solutions, while distributed or hybrid architectures better suit large-scale dynamic scenarios to ensure reliable system operation.

2.5. Summary and Comparison

The evolution of path planning techniques demonstrates a clear methodological trajectory, yet the comparison in

Table 2. Comparison of Path Planning Methods.

Method Categories	Core Mechanism	Planning Scope	Optimality Guarantees	Implementation Complexity	Applicable Scenarios	Representation Methods
Geometric Decomposition	Terrain decomposition with geometric rules; generates parallel rows	Global	Coverage-complete, not distance-optimal	Low	Regular fields with static or sparse obstacles	ox-tracing/bow-shaped path, Partitioning Method
Meta-Heuristics	Population evolution to iteratively optimize paths	Global	Approximate, no guarantee	High	Irregular plots, complex coverage tasks	GA, SA, PSO, ACO
Graph Search	Static Planning: Global search on fixed maps	Global	Globally optimal	Medium	Structured fields, static obstacles	Dijkstra, A*
	Dynamic Replanning: Incremental search for changing environments	Global	Optimal under known information	Medium to High	Dynamic fields with new obstacles	D*, D* Lite
Sampling Optimization	Random sampling to build feasible paths	Global/ Local	Approximately optimal	Medium to High	High-dimensional or complex constraints	RRT, RRT*, PRM
Local Classics	Real-time obstacle avoidance from sensor data	Local	Not guaranteed	Low to Medium	Unknown or dynamic environments	DWA, APF, VFH
Fuzzy Logic	Sensor–rule mapping via fuzzy inference	Local	Not guaranteed	Medium	High-uncertainty environments	FL
Data-Driven	Learning-based mapping from perception to control	Local/ Global	Data-dependent	Very High	Unstructured, highly dynamic settings	RL, DRL (DQN, PPO, DDPG)
Hybrid Methods	Combines global optimality with local reactivity	Global + Local	Global reference; local non-optimal	Very High	Large-scale, dynamic, high-uncertainty fields	CCPP + DWA, A* + APF

reveals that no single approach consistently achieves superior performance across all agricultural scenarios. This reflects the inherent gap between theoretical algorithm design and practical field deployment. Geometric decomposition methods remain extremely efficient and are suitable for structurally regular farmland, but their performance deteriorates sharply in irregular plots or obstacle-rich environments due to their limited adaptability. Graph search algorithms, while capable of ensuring global optimality under static conditions, show limited responsiveness in dynamic environments where real-time adaptation is required. Sampling-based approaches handle high-dimensional constraints effectively but often struggle to ensure smoothness and stability, especially in dense vegetation or cluttered fields.

Meta-heuristic and data-driven methods provide strong global search or adaptive learning capabilities, but their computational cost and parameter sensitivity constrain applicability on embedded agricultural robotic platforms. Classical local planners demonstrate high real-time responsiveness but lack global optimality, making them unsuitable for large-scale field operations when used alone. Multi-robot coordination strategies exhibit promising performance in simulations, yet real-world deployment remains hindered by unstable communication, heterogeneous platform capabilities, and task allocation imbalance. Consequently, practical agricultural robot system design requires flexible integration of multiple planning paradigms, balancing computational feasibility, robustness, and real-time requirements based on environmental complexity, hardware constraints, and operational goals. It should be noted that the performance differences summarized here reflect general trends reported in the literature, rather than controlled experiments under identical hardware or environmental conditions. Therefore, these conclusions are intended for methodological comparison rather than absolute quantitative benchmarking.

2.6. Unified Evaluation Metrics and Experimental Conditions

Although numerous studies have explored path planning and trajectory tracking for agricultural robots, the diversity of evaluation protocols and experimental conditions significantly limits the comparability of reported results. Existing literature frequently adopts different accuracy metrics, computational indicators, hardware platforms, and environmental settings, which can lead to inconsistent interpretations of algorithm performance. To address these challenges and to provide a more rigorous methodological foundation for cross-study comparison, this review introduces a unified evaluation framework and clarifies key experimental factors that influence navigation outcomes.

A comprehensive evaluation of navigation algorithms should first consider geometric accuracy, which remains the core indicator of tracking performance. Metrics such as lateral deviation, heading error, and root-mean-square error directly reflect the robot’s ability to follow a prescribed path. Coverage completeness, overlap rate, non-working distance, and the number of turns further describe operational efficiency, especially in coverage path planning and agricultural field operations requiring repetitive lane-following. Since real-world agricultural robots operate under energy and resource constraints, energy consumption per unit distance, fuel usage, and battery discharge characteristics should also be incorporated into the assessment. Equally important are real-time computational properties, such as control cycle latency, update frequency, CPU and GPU load, and memory footprint, which determine whether a navigation algorithm can be feasibly deployed on embedded hardware platforms. Robustness indicators, including performance degradation under slopes, soil slip, GNSS outages, or sensor noise, provide insight into a method’s stability under realistic disturbances. Finally, environmental descriptors such as field topology, soil moisture, crop density, and illumination conditions should accompany experimental results, as they substantially influence sensing, localization, and control responses.

Despite the importance of these factors, previous studies often report results obtained under widely different hardware and sensor configurations, making direct comparison difficult. Algorithms executed on high-performance onboard computers naturally benefit from lower computation times and smoother control commands, whereas the same algorithms deployed on low-power embedded processors may experience delays or degraded performance. Similarly, variations in GNSS accuracy, LiDAR resolution, camera frame rate, or IMU drift introduce inconsistencies that can mask or exaggerate the true capabilities of a method. Environmental variability further complicates evaluation: soil conditions, terrain slopes, crop occlusion levels, and the presence of dynamic obstacles can significantly alter a robot’s behavior, yet these variables are not consistently documented. Differences in experimental protocols—such as robot speed, trajectory type, obstacle density, or the number of trial repetitions—also affect the reliability of reported metrics, particularly when only best-case results are presented rather than averaged outcomes.

Establishing and adopting a unified evaluation framework, along with transparent documentation of hardware specifications and environmental conditions, is therefore essential for improving scientific rigor in agricultural robot navigation research. By explicitly clarifying the factors that influence algorithmic performance, this subsection provides the methodological basis for interpreting the comparative analyses presented in tables, and ensures that readers can more reliably assess the strengths and limitations of the various planning and control approaches reviewed in this study.

3. Trajectory Tracking Control Methods

Path planning provides agricultural robots with a desired operational trajectory. Theoretically, path planning and trajectory tracking form a closed-loop “Decision-Execution” coupling within the robot’s navigation architecture. As the upper-level decision layer (Section 2), path planning addresses geometric constraints in the Configuration Space to generate continuous paths for spatial guidance. In contrast, trajectory tracking acts as the lower-level execution layer (this section), focusing on dynamic constraints and time-varying disturbances in the State Space [41,42,43]. However, this execution layer faces significant challenges in real-world field operations. Factors such as vehicle dynamic nonlinearity, tire slippage, terrain slopes, crop obstructions, and other environmental uncertainties make trajectory tracking control the critical determinant of final operational precision.

3.1. Modeling and Uncertainties

The prerequisite for designing a high-performance trajectory tracking controller is establishing a mathematical model that accurately describes the motion characteristics of agricultural robots. This model serves as the foundation for controller design, and its precision directly impacts control performance [44,45].

Robot modeling is typically categorized into kinematic and dynamic modeling. Kinematic models describe the relationship between robot pose and velocity inputs based on geometric relationships, disregarding physical parameters such as forces and mass. For wheeled mobile robots, the incomplete constraint model based on the Ackermann steering principle is a common choice. It assumes the robot moves without slip on flat terrain. This model offers high computational efficiency and serves as the foundation for pure tracking, geometric tracking algorithms, and certain model predictive control applications. Dynamic models further incorporate physical factors like forces, torques, mass distribution, and tire-ground interactions, enabling more accurate representation of the system’s nonlinear dynamic response. For tracked platforms, slip and roll effects must be considered; for multi-body systems like tractor-trailer combinations, articulated dynamic models are required. Dynamic models provide the design foundation for advanced algorithms such as model predictive control and sliding mode control, which necessitate system models.

It is particularly noted that trajectory tracking control methods are typically categorized as “model-based (with models)” or “non-model-based (without models),” a distinction primarily reflecting whether mathematical models are explicitly used during controller design. In practice, all control methods implicitly rely on some understanding of the system’s dynamic behavior: PID control implicitly estimates system dynamics through error feedback mechanisms; FL encodes human operational experience into rule bases; RL learns state-action mapping strategies through data interaction. While these methods do not explicitly depend on mathematical analytical models during design, their effectiveness still rests upon some representation of the system’s dynamic characteristics.

Uncertainties faced by agricultural robotic systems primarily stem from internal model mismatches and external environmental disturbances. Internal uncertainties include parameter uncertainties (such as measurement errors in robot mass or moment of inertia, or time-varying parameters caused by changes in tank liquid levels during operations) and structural uncertainties (such as neglected higher-order dynamics, unmodeled nonlinear friction, and actuator delays). External uncertainties encompass variations in gravitational components due to terrain undulations (particularly lateral slopes), fluctuations in tire-ground adhesion coefficients caused by changing soil properties, and time-varying disturbances resulting from operational load changes. He et al. [46] proposed a third-order full-state feedback controller to address dynamic uncertainties in curved path tracking. By improving system performance through pole placement, they achieved a maximum lateral deviation of 0.0705 m and an average absolute error of 0.0218 m on variable-curvature paths—a 31% reduction compared to pre-compensation—effectively enhancing control system stability and accuracy.

In summary, robot modeling provides the theoretical foundation for control system design, while uncertainty remains a critical factor affecting control performance. The various control methods introduced in subsequent sections represent different technical approaches to addressing modeling and uncertainty issues.

3.2. Classical Control Algorithms

Classical control algorithms remain widely adopted in agricultural robots, not because of superior theoretical performance, but due to their transparency, ease of implementation, and predictable behavior under nominal conditions. Their effectiveness, however, is inherently bounded by unmodeled nonlinearities, time-varying disturbances, and coupling effects that are prevalent in real field operations [47].

PID (Proportional, Integral, Differential) control relies on linear error feedback and implicitly assumes quasi-linear system behavior. While acceptable for low-speed straight-line tracking on flat terrain, fixed-gain PID controllers exhibit pronounced performance degradation under curvature transitions, soil-induced slip, or payload variations. Retuning gains for one operating condition often leads to instability or sluggish response in others, limiting its applicability to narrowly defined scenarios [48].

Geometric tracking methods such as Pure Pursuit Control (PPC) and the Stanley controller bypass explicit dynamic modeling by exploiting geometric relationships between the vehicle and the reference path. This simplicity enables stable and smooth control at low computational cost. However, their reliance on kinematic assumptions makes them sensitive to speed variations and curvature changes [49]. In particular, the look-ahead distance in PPC introduces a fundamental trade-off between responsiveness and stability, while the Stanley method may induce aggressive steering under high lateral error, even at moderate speeds [50]. Ahn et al. [51] proposed a heuristic method for adjusting the pre-aim point. By dynamically selecting the pre-aim point, they significantly improved the vehicle’s tracking accuracy and stability on high-curvature paths, reducing the cross-track error by 41% compared to traditional PPC.

As a result, classical controllers should be viewed as baseline solutions or components within hybrid control architectures, rather than standalone solutions for high-precision agricultural navigation. Their continued relevance lies in providing reliable fallback behavior and interpretable control actions, especially when advanced model-based or learning-based controllers encounter uncertainty or computational limitations.

3.3. Advanced Control Algorithms

To overcome the limitations of classical algorithms and enhance tracking accuracy and robustness against uncertainty, a series of modern advanced control techniques have been introduced into trajectory tracking for agricultural robots.

Linear Quadratic Regulator (LQR) is an optimal control method based on state feedback. It solves for optimal control laws for linear systems by minimizing a quadratic cost function that incorporates both state error and control energy consumption. This method exhibits high theoretical completeness, providing analytical solutions that guarantee system stability. The pronounced nonlinearity and time-varying characteristics of agricultural robot dynamics limit the direct application of standard LQR. Nevertheless, LQR remains valuable in practical contexts through multiple approaches: designing local controllers by linearizing systems near equilibrium points; serving as the underlying solver for Model Predictive Control; or functioning as a benchmark for evaluating other controller performance. LQR retains significant importance in theoretical framework development and understanding advanced control methodologies. Amertet et al. [52] tested the LQR control system for agricultural robots, determining that the peak, settling, rise, and steady-state times for the low-cost control were 0.0707 s, 5.53 s, 3.11 s, and 0.0707 s, respectively. Despite its elegant theoretical properties, the direct applicability of LQR in agricultural robots is inherently constrained by model linearization errors and unmodeled disturbances, which often dominate system behavior in real field conditions.

Model Predictive Control (MPC) stands as one of the most mainstream and extensively researched advanced methods today. Its core principle involves rolling optimization and feedback correction, implemented through three key phases: first, establishing a prediction model (e.g., kinematic for harvesting tasks or dynamic for high-speed spraying); second, formulating the objective function JJ to minimize state deviation and control increments via weighting matrices Q and R: $J={\sum _{k=1}^{N_p}‖x(t+k|t)-x_{ref}‖}_Q^2+{\sum _{k=0}^{N_c-1}‖\Delta u(t+k|t)‖}_R^2$; and finally, executing the receding horizon strategy. At each control cycle, MPC utilizes an embedded predictive model to forecast the system’s state evolution over a future time interval based on the current vehicle state. It then solves this optimization problem (where the cost function typically includes tracking error, control magnitude, control rate of change, etc., subject to various constraints such as steering angle limits, speed limits, and stability constraints) to generate a sequence of optimal control commands for the future time interval [53,54]. However, only the first control command is applied to the system. This process repeats in the next cycle. MPC explicitly handles the system’s multi-input multi-output characteristics, diverse constraints, and future dynamic responses, thereby generating exceptionally smooth and precise control commands. Parameter selection and optimization are critical in agricultural applications: the prediction horizon N_p must be sufficient to cover the vehicle’s response time during headland turns, while the sampling time T_s is typically set between 50 and 100 ms to filter high-frequency vibrations from uneven terrain. To address solver failure caused by slippery soils and actuator saturation, an optimization scheme relaxes hard constraints into soft constraints using slack variables, allowing slight control violations during momentary wheel slip to ensure continuous controller operation and system robustness.

However, MPC’s high computational complexity demands significant processing power for solving optimization problems online, potentially making real-time operation challenging on low-power embedded agricultural machinery platforms. Extensive research focuses on reducing MPC computational burden or enhancing its performance. For instance, Han et al. [55] investigated learning-based MPC, utilizing Gaussian process regression to learn and compensate for model errors in real time, reducing average error by 23.64% compared to traditional MPC. In agricultural robots, this limitation is further amplified by cost-sensitive hardware selection and harsh operating conditions, where high-performance onboard computation is often unavailable or unreliable over long-term deployment.

Sliding Mode Control (SMC) is another advanced control method renowned for its strong robustness. The design of SMC involves two steps: first, designing a sliding surface that is a function of the system state, ensuring stable and favorable dynamic performance when the system state is constrained on this surface; second, designing a control law that drives the system state to reach the sliding surface within a finite time and maintain it there. SMC is highly suitable for systems with unknown disturbances and parameter variations, such as slopes or wet, muddy environments [56,57]. More importantly, in agricultural machinery with hydraulic or electric steering actuators, excessive chattering may accelerate mechanical wear and increase energy consumption, which constrains the long-term applicability of pure SMC in continuous field operations. To suppress chattering, advanced sliding mode techniques—including higher-order sliding modes (super-helix algorithms), fuzzy sliding mode control, and boundary layer methods—have been extensively studied. Liu et al. [58] tested a parameter-predictive super-helix SMC algorithm, demonstrating that at speeds of 0.5 and 1.0 m/s, this method reduced absolute mean deviation by 69.2% and 50%, respectively, compared to traditional SMC, while reducing heading deviation by 61.1% and 40%.

Adaptive Disturbance Rejection Control (ADRC) and its core component, the Expanded State Observer (ESO), constitute another powerful tool for handling uncertainty. The ESO approach unifies internal model mismatches and external disturbances into a single aggregate disturbance, expanding it into a new state variable for real-time estimation via an observer design. Once estimated, this aggregate disturbance is feedforward-compensated within the control law, actively transforming the uncertain system into a simple series integrator configuration for easier control. ADRC does not rely on precise mathematical models and can estimate and compensate for disturbances in real time, significantly improving tracking accuracy on slopes, wet soil, and similar conditions. Sun et al. [59] designed fixed-time nonsingular terminal sliding mode control based on adaptive disturbance observers for unmanned agricultural machinery. This method not only enables rapid and accurate estimation and compensation of unknown disturbances but also guarantees convergence of tracking error within a fixed time. Huang et al. [60] proposed a fusion architecture combining ESO and SMC for skid estimation and nonlinear disturbance compensation. This approach maintained lateral error standard deviation < 0.04 m even under skid-enhanced conditions. However, observer parameter sensitivity significantly impacts performance, necessitating robust parameter selection strategies for engineering deployment. Nevertheless, the performance of ADRC-based approaches is highly sensitive to observer bandwidth and tuning heuristics, and improper parameter selection may introduce noise amplification or phase lag, particularly under high-frequency disturbances induced by rough terrain.

3.4. Intelligent and Adaptive Control Algorithms

As discussed in Section 2.2.2 and Section 2.2.3, intelligent algorithms such as reinforcement learning and fuzzy logic have demonstrated significant potential in the field of path planning. Similarly, at the level of trajectory tracking control, these data-driven and knowledge-driven approaches offer novel perspectives for addressing system nonlinearity and uncertainty challenges, differing from traditional modeling approaches. This section will focus on how these methods can be directly applied to generate control commands or optimize controller parameters, thereby achieving higher levels of adaptive and robust performance.

Adaptive Control (AC) is an advanced class of methods specifically designed for online, real-time adjustment of controller parameters in response to changes in system dynamics [61]. When the mathematical model parameters of the controlled object are unknown or undergo significant changes, adaptive laws adjust controller parameters in real time, enabling the system output to track desired performance metrics or reference models. Adaptive control is primarily divided into two major categories: Model reference adaptive control (MRAC) designs an ideal reference model and parameter adjustment mechanism to achieve asymptotic tracking of the reference output by the system output. Self-Tuning Control (STC) achieves performance optimization by identifying system parameters online and updating the controller design in real time. In agricultural robotics, operational load variations and changes in ground mechanics cause gradual or abrupt shifts in vehicle mass, center of gravity position, and tire-ground interaction parameters. Traditional fixed-parameter controllers experience significant performance degradation under such conditions. Adaptive controllers automatically compensate for these parameter changes, maintaining system tracking accuracy and stability without requiring manual parameter re-tuning. Adaptive control (AC) possesses robust capabilities in handling parameter uncertainties, making it a key technological approach to enhance the robustness of agricultural robots under variable load and terrain conditions.

Neural Networks (NN), leveraging their powerful nonlinear fitting capabilities, are widely employed to compensate for system modeling errors or dynamically challenging characteristics. A common approach involves NN-assisted control: maintaining MPC or SMC as the primary controller framework while using NN to learn and compensate for uncertainties in the dynamic model online, thereby enhancing control accuracy and robustness. Another approach is direct neural adaptive control, where specific network architectures and adaptive laws are designed to enable online updating of network weights, directly outputting control commands.

RL learns optimal control strategies through trial-and-error interactions with the environment. In trajectory tracking problems, the controller observes environmental states (such as tracking error and vehicle posture), executes corresponding actions, and continuously updates the strategy based on a reward function, ultimately forming an optimal mapping from states to control actions. DRL further incorporates deep neural networks as function approximators, enabling it to process high-dimensional state inputs like camera images. RL’s prominent advantage lies in its exceptional environmental adaptability, making it particularly suitable for agricultural machinery operations in dynamic settings. However, RL training often requires substantial data or simulation interactions, may exhibit unstable training processes, and carries risks of strategy transferability in real-world environments. Zheng et al. [62] proposed a GPR-based learning-augmented MPC (LB-MPC), achieving an average lateral error reduction of 25% under model uncertainty. By employing warm-starting techniques, they reduced online optimization time from 0.12 s to 0.04 s, demonstrating the engineering potential of integrating learning models with predictive control, while still requiring rigorous safety envelopes and fallback controllers for field deployment.

FL excels at handling imprecise and ambiguous information. In trajectory tracking control, fuzzy controllers quantify tracking errors and their rates of change as fuzzy variables. These are then reasoned through a rule base constructed from expert experience before being defuzzified into precise control commands. Fuzzy control requires no precise mathematical models and exhibits strong robustness, making it particularly suitable for scenarios where experienced operators can clearly describe control rules, but precise equations are difficult to establish. Liu et al. [63] investigated fuzzy predictive function control for path tracking in rice transplanters based on feedback linearization. While ensuring the real-time performance of the path tracking system, this approach enhanced control accuracy and improved dynamic performance.

Overall, intelligent and learning-based methods can integrate with traditional or modern control approaches to form hybrid architectures. This enables simultaneous achievement of performance, robustness, and adaptability, thereby realizing high-precision and high-reliability control in complex agricultural environments.

3.5. Summary and Comparison

The evolution of trajectory tracking control methods reveals a clear transition from model-independent to model-dependent strategies, from linear feedback to nonlinear compensation, and from manually tuned parameters to adaptive or data-driven learning schemes. As summarized in

Table 3. Comparison of Trajectory Tracking Control Methods.

Category	Control Algorithms	Core Mechanism	Model Dependency	Computational Complexity	Applicable Scenarios
Classic Algorithms	PID	Linear error-based feedback	None	very low	Low-speed linear tracking; moderate-precision tasks
	PPC	Kinematic forward-point steering	Weak (Kinematic Only)	low	Moderate-speed tracking with gentle curvature
	Stanley	Nonlinear heading–lateral error fusion	Weak (Kinematic Only)	low	Ground vehicles; mid–high speed tracking
Advanced Algorithms	LQR	Optimal linear control minimizing quadratic cost	Strong (Precise linear model)	Offline: high; Online: low	Structured, flat environments
	MPC	Predictive optimization with rolling horizon	Strong	high	High-precision field tracking; slopes
	SMC	Nonlinear Sliding manifold tracking under disturbances	Medium	Medium	High disturbance environments, slopes
	ADRC	ESO-based active and nonlinear disturbance rejection	Weak	Medium	Slippery soil; varying loads
Intelligent Adaptive Algorithms	AC	Online controller adjustment via nonlinear parameter adaptation	Medium	Medium	Time-varying dynamics (load, inertia)
	NN	Nonlinear Neural network model approximation	None	Inference: Medium; Training: extremely high	Highly nonlinear, difficult-to-model systems
	RL/DRL	Learning optimal linear/nonlinear strategies via interaction	None	Inference: Medium; Training: extremely high	Dynamic, unstructured environments
	FL	Fuzzy rule-based nonlinear control	None	Medium	High-uncertainty real-time control

, classic controllers such as PID, PPC, and Stanley offer low computational cost and can achieve sub-decimeter accuracy (<0.1 m) in low-speed and structured environments. However, their performance degrades substantially under high-curvature, high-speed, or disturbance-rich conditions due to limited robustness.

Advanced methods significantly enhance tracking performance, typically reducing lateral deviation to 2–5 cm while providing explicit constraint handling and robustness against model uncertainties. Their main limitations lie in increased computational demand (often 5–20 ms per iteration for MPC on embedded platforms), sensitivity to model accuracy, and the need for sophisticated parameter tuning.

Intelligent and adaptive control methods further enhance adaptability under nonlinear, time-varying, or poorly modeled dynamics. Nevertheless, most such methods remain at the simulation stage, lacking extensive field validation and exhibiting high training or tuning overhead. In practical deployments, hybrid control structures have become mainstream, for instance, combining MPC for upper-layer trajectory generation with PID or fuzzy control for low-level actuation; or integrating extended state observers with SMC/MPC to enhance disturbance rejection.

Overall, selecting an appropriate tracking controller depends on operational constraints, required accuracy, and available onboard computation. The comparison in Table 3 highlights the trade-offs among robustness, computational capacity, and implementation complexity across different control paradigms.

To provide a more application-oriented complement to the methodological comparison presented in Table 2 and Table 3,

Table 4. Summary of Research Findings on Selected Path Planning and Trajectory Tracking Methods.

Author (Year)	Main Methods	Lateral Error	Coverage	Calculation Time	Field Trials	Advantages	Disadvantages
He (2023) [64]	Dynamic Head-to-Ground Turn Optimization	0.02 m	-	-	Yes	Enhanced head-to-ground efficiency and stability	Needs adaptation for complex obstacles
Ji (2023) [65]	Adaptive SMC	Lowest MAE/RMS	-	-	No	High tracking accuracy; reduced chattering	Simulation only; needs real-world validation
Zhu (2023) [66]	FL + improved Beetle Antennae Search	Lower mean steady-state error than PPC	-	Offline optimization reduces real-time computation load	Yes	Higher accuracy and stability; avoids real-time performance issues via offline optimization	Performance depends on the accuracy of the offline simulation model
Li (2023) [67]	SMC with a nonlinear observer	Avg: 0.0247 m	-	-	Yes	robust tracking in paddy fields with slip and disturbances; no overshooting	Assumes standardized rectangular farmlands
Chen (2025) [68]	a 3D fuzzy controller for dynamic preview distance and a feedback compensator	Standard deviation of lateral offset reduced by 11.3% to 40.4%	-	-	Yes	Improves navigation accuracy and stability; robust against skidding and off-track driving	Not applicable in scenarios where a satellite navigation baseline cannot be established
Santos (2020) [69]	extended A*	-	-	0.24–0.26 s	Yes	Generates safer paths on steep slopes to avoid dangerous postures	Path length may be longer than standard A*; memory usage needs optimization for larger areas
Zhang (2025) [70]	Segmented Bézier curves in a convex QP problem, with MPC and PID	Max: 0.0400 m	-	50–200 ms	Yes	Generates high-order continuous and feasible trajectories for multi-task scenarios; ensures high-precision tracking	Primarily targets static environments without integrating real-time perception
Peng (2024) [71]	pattern-based or Hybrid A*	<0.1 m	-	8.2–12.3 s	Yes	Generates smooth, collision-free, kinematically feasible trajectories in constrained, irregular headlands with obstacles	Slower than classic pattern-based planners; may fail in extremely narrow spaces under time limits
Monsalve (2022) [72]	PID	Small line error	-	-	No	Low-cost system using fewer sensors (odometry-based) than vision-based alternatives	Performance at turning points needs improvement; susceptible to wheel slip
Xie (2024) [73]	Visual navigation + improved LQR	Avg: 0.0102 m	-	>38 fps for detection	Yes	Good robustness and accuracy	Cannot autonomously perform U-turns
Cao (2025) [74]	Improved RRT*	-	-	0.357–0.565 s	Yes	Robust mapping and localization in complex orchards; plans safe and smooth paths	High computational demand; difficult for long-term, large-scale mapping
Sánchez (2025) [75]	Double spiraliform path planning with SMC	Avg: <0.012 m	generates double spiraliform coverage paths	-	No	Generates efficient paths for complex fields	Chattering is present; idealized simulation without real-world validation
Sun (2023) [76]	Fuzzy Adaptive Recursive Terminal SMC	RMSE for circular path: <0.046 m	-	-	Yes	High tracking precision and strong robustness; chattering is alleviated by the fuzzy system	Dead-reckoning errors can accumulate over long trajectories
Willekens (2025) [77]	PPC + PID	0.01 m	-	-	Yes	Open-source, high accuracy, multi-robot support	Complex, requires specific hardware
Fujinaga (2025) [78]	Hybrid waypoint + cultivation bed navigation using LiDAR	±0.05 m	94.73%	-	Yes	No infrastructure needed, robust in dynamic environments	No obstacle handling, LiDAR affected by sunlight, static map
Yang (2022) [79]	Optimal goal point algorithm for path tracking	Avg: 0.052 m; Max: 0.061 m	-	-	Yes	Improves tracking accuracy, adaptive look-ahead point	Not specified

Note: The quantitative ranges presented in this table summarize typical values reported across representative studies and may vary depending on hardware platforms, sensor configurations, environmental conditions, and evaluation methodologies.

summarizes representative empirical studies on agricultural robot navigation. Unlike the previous tables, which focus on algorithmic characteristics and theoretical capabilities, Table 4 emphasizes real-world performance indicators, such as lateral deviation, coverage rate, turning efficiency, computation time, and whether field trials were conducted. By presenting these metrics side-by-side, Table 4 bridges the gap between algorithm design and practical implementation, revealing how different approaches behave under actual agricultural conditions. All numerical results in Table 4 are directly extracted from the original publications, and units remain consistent with the cited studies to avoid introducing cross-paper normalization bias.

The results show that path planning studies often highlight improvements in coverage completeness or reductions in non-working distance, whereas trajectory tracking studies typically report centimeter-level accuracy and enhanced robustness achieved through advanced control techniques such as MPC, SMC, and disturbance-observer-based methods. Meanwhile, intelligent and learning-based methods demonstrate promising adaptability but still face challenges related to training overhead and generalization across different field environments.

Overall, the comparative evidence presented in Table 4 reinforces the observations made in Section 2.6: differences in hardware platforms, sensor configurations, environmental complexity, and experimental protocols substantially influence reported performance. When interpreted within the unified evaluation framework established earlier, the results highlight both the strengths and current limitations of existing navigation approaches. While many methods achieve high accuracy or improved efficiency under specific conditions, their scalability and engineering feasibility remain constrained in complex or large-scale agricultural settings. These insights underscore the importance of standardized benchmarks and comprehensive field validation. The following section examines typical agricultural environments—such as open-fields, orchards, greenhouses, and hilly slopes—to further contextualize the practical challenges faced during real-world deployment.

4. Applications and Challenges

The summary and comparison of path planning and trajectory tracking control methods in Section 2 and Section 3 reveal that different algorithms possess distinct advantages in terms of accuracy, efficiency, and robustness. However, their performance is highly dependent on the specific operational environment. Agricultural scenarios vary significantly in spatial structure, environmental disturbances, agronomic requirements, and topographical conditions. These differences directly dictate the selection and adaptation strategies for perception, positioning, planning, and control technology modules. A navigation system that excels in open fields may prove completely ineffective in dense orchards or rugged slopes. Therefore, thoroughly analyzing the characteristics of different typical application scenarios and conducting targeted technical adaptation and innovation based on these insights is crucial for advancing agricultural robots from the laboratory to the field. This chapter will focus on examining the current application status and unique challenges in four typical scenarios: open-fields, orchards, greenhouses, and hilly slopes. It will further summarize the common technical bottlenecks currently facing the entire field.

4.1. Typical Application Scenarios

4.1.1. Open-Fields

Open-fields typically refer to large-scale, relatively flat agricultural land with uniformly cultivated crops (such as wheat, rice, or corn). This represents the most widespread and mature application scenario for autonomous agricultural driving technology. The open space, regular field boundaries, and minimal obstructions in open fields enable GNSS/RTK technology to provide stable and reliable centimeter-level absolute positioning information. Through tight or loose coupling fusion with IMU, GNSS/INS integrated navigation systems can further compensate for positioning errors during short-term GNSS signal loss and provide vehicles with complete attitude information. Consequently, navigation systems in large-scale farmland environments exhibit relatively low dependency on environmental perception sensors like vision or LiDAR.

For path planning, CCPP remains the most widely applied method. Given the regularity of field plots, traditional geometric approaches retain their status as the most common and economical choice due to their exceptional computational efficiency and ease of implementation. However, for larger fields or those containing scattered obstacles, purely geometric methods may lead to reduced efficiency. In such cases, intelligent optimization algorithms like GA or PSO can be employed to optimize the traversal sequence of work units or headland turning strategies, significantly reducing non-working path length and total operation time. Global point-to-point planning is commonly used for field-to-field transitions in large-scale operations, with the A* algorithm frequently chosen for its efficiency on regular grid maps. Regarding control, field operation paths primarily consist of long straight lines supplemented by regular headland turns. MPC demonstrates significant advantages in this scenario, explicitly handling vehicle dynamics constraints to generate exceptionally smooth control commands, particularly achieving high precision in straight-line tracking. The PPC algorithm remains simple and reliable, widely adopted in many low-cost autonomous systems, with performance optimized through adaptive strategies like adjusting the look-ahead distance. However, field environments are not free from disturbances. Variations in soil moisture can cause wheel slippage, introducing lateral tracking errors. The disturbance observer ESO can be integrated into control strategies to estimate and compensate for slip disturbances online, thereby maintaining tracking accuracy. Numerous scholars have tested relevant navigation algorithms in field scenarios:

Cui et al. [80] designed a navigation control system for field management robots and conducted multiple field trials. Results demonstrated that the system’s trajectory error met engineering thresholds during routine field operations (seeding, pesticide application) while maintaining stable performance. They emphasized the critical role of system integration for engineering success, though sufficient data on long-term robustness under extreme weather/muddy conditions remains unavailable. Zhou et al. [81] proposed an improved MPC scheme combining genetic algorithms with adaptive time-domain parameters for complex curved path tracking on articulated tractors. Simulations and field tests showed maximum lateral deviations reduced by 59.0%, 24.9%, and 13.2% for U-shaped/figure-eight/complex curves, respectively. Field validation showed corresponding maximum deviations reduced by 67.8%, 44.7%, and 45.1% compared to traditional MPC. This method significantly enhances transient and steady-state performance during field curve operations but relies heavily on model accuracy and online computational resources. Li et al. [67] proposed a robust tracking scheme combining a nonlinear observer with fast terminal sliding mode control for skid-prone rice transplanting/paddy field scenarios. It maintains stable tracking under wet/muddy conditions. At an operating speed of 1.0 m/s, the average absolute lateral deviation during straight-line tracking was 0.0247 m, with a mean squared error of 0.0311 m, demonstrating significantly lower lateral error than uncompensated control.

Technical bottlenecks in field navigation primarily include: (1) Despite open terrain, GNSS signal attenuation due to multipath effects or extreme weather remains a concern; (2) Systems must possess reliable real-time obstacle avoidance capabilities to respond to sudden moving obstacles; (3) Efficiency optimization for large-scale operations: Further refining path planning and scheduling for multi-robot coordination to maximize overall farm system productivity.

From a scenario-driven perspective, open-field environments favor navigation methods that emphasize global optimality, coverage efficiency, and computational simplicity. Structured coverage path planning combined with kinematic controllers, such as PPC or simplified MPC, remains the most practical choice under GNSS-available and geometrically regular conditions. In contrast, learning-based local planners or highly complex control schemes often provide limited additional benefit in such scenarios while increasing system complexity and tuning difficulty. Therefore, for large-scale open-field operations, robustness, scalability, and interpretability should be prioritized over algorithmic sophistication.

4.1.2. Orchards

Orchards present entirely different challenges for agricultural robot navigation. Dense shading and complex inter-row structures severely obstruct satellite signals, causing GNSS positioning accuracy to plummet or fail entirely, rendering it unreliable as a sole positioning source [82]. While crop rows exhibit some structural regularity, they often feature irregularities, breaks, or gaps. Tree trunks in the environment serve as both static obstacles requiring avoidance and natural beacons for positioning. The technological core of orchard navigation shifts from reliance on GNSS to multi-sensor fusion and perception of natural environmental features. Simultaneous Localization and Mapping (SLAM) technology is employed, utilizing LiDAR or visual features to construct point cloud maps or semantic maps of the orchard environment, enabling precise positioning [83,84,85].

Path planning must be tightly integrated with perception. Semantic maps become critical, as planners need to identify “where crop rows are” and “where tree trunks are” to generate navigation paths that meet agronomic requirements while actively avoiding obstacles. Regarding tracking control, orchards demand extremely stringent lateral tracking accuracy; even centimeter-level deviations can cause robots to collide with tree trunks, damaging fruit trees or the robot itself. PPC algorithms remain applicable here but exhibit limited performance in curved rows. Advanced control methods like high-precision SMC or MPC combined with disturbance observers ensure stability and accuracy on complex paths. Liu et al. [25] proposed a comprehensive BL-ACO + GO-SMC control scheme for orchard wheeled mowers. Field tests demonstrated a 47.6% reduction in operation time and fuel consumption, alongside significant improvements in tracking time and energy efficiency within narrow aisles. This highlights the engineering value of integrating cooperative optimization and control for orchard tasks, though further scenario validation is needed for generalization to dense obstacles and uneven terrain. Raikwar et al.’s [86] model-based orchard navigation design achieved normalized root mean square errors of steering deviation between 0.2 and 0.4 degrees during turns and 0.05 degrees during straight paths. This demonstrates maintainable accuracy without GNSS by fusing wheel speed/steering encoders with vision/odometer data, though it faces limitations under prolonged drift conditions. Li et al.’s [84] path tracking and anti-slip control scheme for horticultural mowers confined lateral error to 0.05 m and longitudinal error to 0.04 m under orchard slippage conditions. This demonstrates that orchard-specific anti-slip control combined with MPC cascade control is highly effective in narrow rows and high-slip environments.

Key challenges in orchard navigation include: (1) Extreme demands on perception algorithm robustness due to light variations, foliage occlusion, and seasonal appearance changes; (2) Addressing environmental alterations during prolonged operation (e.g., fruit harvesting, pruning); (3) Developing higher-precision control algorithms and actuators.

In orchard and row-based agricultural scenarios, navigation performance is strongly constrained by narrow passages, dense vegetation, and frequent occlusions. Under these conditions, reliance on global planning alone is insufficient, and methods integrating local perception and reactive planning become essential. Controllers with stronger disturbance rejection capabilities, such as MPC or robust control variants, are generally more suitable than purely geometric approaches. However, excessive model complexity may reduce reliability in long-term deployment. Consequently, hybrid architectures that combine lightweight global planning with perception-driven local control represent a more balanced and effective solution.

4.1.3. Greenhouses

The interior space of greenhouses is confined, with facilities such as cultivation beds, pipes, and support frames densely packed. Passageways are typically fixed and narrow. Lighting conditions may vary due to shading from the greenhouse roof and supplemental lighting. The high-temperature, high-humidity environment poses challenges to the reliability of sensors and electronic equipment [87]. Operational tasks are also intricate and diverse, including pot-by-pot inspections, precision irrigation, and fruit harvesting. Greenhouse navigation relies entirely on indoor perception technologies. Given the highly structured environment, artificial beacons like QR codes or ArUco markers are widely used to provide absolute pose information. They offer low cost and high reliability but require pre-deployment. Autonomous positioning solutions based on visual odometry or laser SLAM are increasingly adopted, though they must address challenges posed by repetitive textures (uniformly arranged crops) and dynamic changes (plant growth, personnel movement). Ultrasonic and infrared sensors can be employed for close-range obstacle avoidance.

Greenhouse paths are typically predefined fixed routes, reducing the importance of path planning while emphasizing real-time obstacle avoidance and precise docking. Precise motion control is critical due to typically low operating speeds and high demands for absolute positioning accuracy. PID control and pure tracking algorithms often suffice in such low-speed, structured environments. More complex tasks require simple multi-robot scheduling strategies to prevent deadlocks within aisles. Kulathunga et al. [88] proposed a navigation framework for narrow passages in strawberry tunnels/greenhouses, incorporating point cloud processing and constrained-space trajectory refinement. Field tests achieved an average lateral deviation of 0.08 ± 0.01 m, outperforming several baselines. This demonstrates the effectiveness of point cloud-driven trajectory refinement in highly occluded, narrow-spaced environments. However, the computational and perception pipeline demands high-performance, low-latency hardware. Kang et al. [36] combined DRL with path points for greenhouse navigation in confined spaces, achieving significantly improved trajectory accuracy compared to traditional methods. However, training and domain adaptation incur high costs, and robustness under variable lighting/humidity conditions in real greenhouses requires further validation.

Key challenges in greenhouse navigation include: (1) achieving reliable, low-cost localization without extensive manual labeling; (2) handling frequent interactions with staff and other mobile robots; (3) system integration and cost control: greenhouse robots demand exceptional cost-performance ratios, requiring technical solutions that balance functionality and affordability.

Greenhouse environments impose strict constraints on localization reliability, maneuverability, and safety due to GNSS denial, confined spaces, and structured but cluttered layouts. Navigation methods in this context must prioritize real-time responsiveness and robustness to sensor noise. Reactive planners and local trajectory optimization methods demonstrate higher adaptability than pre-defined global paths. From a control perspective, low-speed precision control with explicit constraint handling is more critical than aggressive tracking performance. As a result, simplicity and reliability outweigh global optimality in greenhouse navigation system design.

4.1.4. Hilly Slopes

Hilly slopes are commonly found in tea plantations or mountainous agricultural areas, with continuously varying gradients as their defining characteristic. Slope inclination significantly alters a robot’s dynamic properties, rendering flat-ground models ineffective and introducing constant gradient perturbations. GNSS signals may become unstable due to terrain obstruction. Additionally, slope operations carry rollover risks, making safety a primary consideration. The ground may be slippery or unstable, exacerbating tire slippage. When planning paths on slopes, optimization goals extend beyond coverage and path length; incorporating stability and energy consumption into the cost function is crucial. Planning algorithms require integration of terrain information such as digital elevation models or 3D point cloud maps. Pour Arab et al. [89] proposed incorporating slope stability metrics into path planning cost functions to automatically avoid excessively steep hazardous areas. Another strategy involves planning based on energy consumption models, selecting paths with gentler slope changes and lower fuel or power consumption to enhance economy while ensuring safety.

At the control level, methods like PID or pure tracking exhibit significant performance degradation on slopes, as they cannot handle continuous disturbances caused by inclines. Robust control and disturbance compensation techniques can address slope disturbances. For instance, incorporating a slope disturbance term into SMC design; or employing ESO to estimate and compensate slope disturbances in real-time as part of the overall disturbance. Xu et al. [90] proposed a dynamic model incorporating road slope disturbances and an adaptive MPC method for lateral slope farmland. By preemptively compensating for slope effects, they achieved a 34% reduction in maximum lateral error magnitude and a 41% reduction in maximum heading deviation magnitude when the lateral slope angle varied slowly and continuously within a 10° range, significantly improving tracking performance. Song et al. [91] proposed an adaptive forward-looking distance tracking control method based on Sparrow Search Algorithm (SSA) for slope skidding issues. Field tests demonstrated stable lateral deviations averaging ≤0.030 m and maximum deviations ≤ 0.106 m during slope operations, indicating significant compensation for skidding and dynamic constraints via adaptive forward-looking distance. However, under extreme slippery conditions, it still requires collaboration with a slip estimator. Simultaneously, control systems must integrate inclinometers to monitor vehicle posture in real time, triggering safety measures like deceleration or stopping when approaching stability limits. Zhao et al. [92] proposed a trajectory prediction method based on slip parameter estimation, with field tests demonstrating significantly improved prediction accuracy for tracked robots navigating muddy slopes, providing reliable prior information for path tracking in sloped terrain.

The challenges of slope navigation are exceptionally severe: (1) Acquiring and maintaining high-precision field topography data is a fundamental challenge; (2) Control stability under strong disturbances: Developing robust controllers capable of resisting continuous and time-varying slope disturbances; (3) Establishing a comprehensive safety evaluation system and emergency response strategies to prevent severe accidents such as rollovers.

Sloped and unstructured agricultural terrains introduce significant dynamic uncertainties arising from soil deformation, wheel slip, and load variation. In such scenarios, navigation methods based solely on kinematic assumptions often exhibit degraded performance. Planning and control strategies that explicitly account for vehicle dynamics, external disturbances, or adaptive parameter variation become more appropriate. Nevertheless, increased model fidelity must be balanced against sensing and computational limitations. Therefore, robust or adaptive control frameworks combined with conservative planning strategies currently offer the most feasible approach for navigation on challenging terrains.

4.2. Common Challenges Currently Faced

Despite the unique challenges posed by different application scenarios, the overall development of agricultural robot navigation technology still faces a series of common technical bottlenecks that constrain its large-scale commercial application.

(1)

Generality and robustness of environmental perception and understanding. Current perception algorithms are predominantly trained on specific datasets, exhibiting severely limited generalization capabilities across different crops, growth stages, weather/lighting conditions, and regional field types. For instance, a crop row detection model trained on early-summer wheat fields may experience drastic performance degradation in late-autumn cornfields or foggy dawn conditions. Developing perception models with strong generalization capabilities represents a core challenge in achieving agricultural robot universality;

(2)

Cross-scenario adaptability of navigation systems. Existing systems are largely tailored for specific environments. A navigation system designed for open fields cannot directly operate in orchards or on slopes. Developing adaptive frameworks that automatically recognize scene changes and dynamically adjust perception, planning, and control parameters is essential for advancing intelligent agricultural robots. This requires algorithms with online learning and adaptive adjustment capabilities;

(3)

The “Sim-to-Real” gap and distribution shift. The phenomenon of “simulation excellence but field failure” fundamentally stems from the distributional shift between training and deployment domains. Root Cause Analysis: First, perception systems struggle with illumination dynamics; while simulators typically employ uniform lighting, real-world high-contrast shadows at noon are often misclassified by vision systems as obstacles (False Positives), causing frequent emergency stops. Second, control policies are limited by simplified contact physics; standard simulators use rigid-body Coulomb friction, ignoring soft-soil sinkage and bulldozing effects, which leads to theoretically aggressive strategies causing immobilization in real mud. Empirical Data Analysis indicates that End-to-End Deep Reinforcement Learning (DRL) agents trained in static environments suffer a performance drop of over 40% in lateral accuracy when facing unmodeled wind gusts or uneven terrain, with error rates spiking nonlinearly under variable lighting. Improvement Strategies: To bridge this gap, frameworks must shift from “open-loop training” to “closed-loop adaptation.” Key measures include: (i) Implementing Domain Randomization during training to apply broad perturbations to lighting textures and friction coefficients, thereby expanding the policy’s robustness boundary; (ii) Utilizing CycleGAN (Cycle-Consistent Generative Adversarial Network) -based style transfer to align simulation imagery with realistic field visual features; (iii) Deploying Online Meta-Learning modules that fine-tune network weights in real-time using actual field interaction data, enabling intelligent algorithms to continuously evolve within unstructured reality;

(4)

Balancing computational complexity and real-time performance. Many advanced algorithms demand substantial computational resources, yet agricultural robots are typically battery-powered with limited computational capacity and power budgets on onboard platforms. Streamlining, optimizing, and hardware-accelerating these algorithms to meet real-time requirements on low-power embedded platforms represents a critical engineering challenge;

(5)

Practical implementation of multi-robot coordination systems. While multi-robot coordination theoretically offers substantial efficiency gains, its actual deployment faces significant hurdles. Communication latency, bandwidth constraints, and reliability issues become particularly pronounced in large-scale farmland. Task allocation and path planning among agricultural robots of different models and functionalities further complicate coordination. Additionally, the absence of unified standards and communication protocols hinders interoperability between devices from different manufacturers;

(6)

Lack of long-term reliability validation and cross-seasonal performance assessment. Current validation frameworks suffer from the limitations of “snapshot” evaluations, relying predominantly on single-instance, short-duration experimental data that fail to reveal cumulative errors and system fatigue during extended operations. Agricultural environments exhibit significant long-term time-varying characteristics: the transition across growth cycles (from bare soil at sowing to full canopy at harvest) causes drastic drifts in visual feature distribution; prolonged continuous operation leads to sensor dust accumulation, lens fogging, and thermal noise, which significantly degrade perception confidence; and dynamic changes in soil moisture (alternating between dry compaction and post-rain mud) fundamentally alter the tire-terrain adhesion model. Therefore, validation standards must shift from simple “tracking accuracy” to lifecycle “long-term autonomy” assessment. Future research must incorporate longitudinal comparative experiments across seasons and adopt “Mean Time Between Interventions” and “Failure Recovery Rate under Extreme Weather (glare/heavy rain)” as core metrics for engineering maturity, demonstrating the continuous reliability of algorithms throughout the real-world agricultural production cycle;

(7)

Cost–Benefit Tradeoffs. Ultimately, any technology must pass the test of economic viability for widespread adoption. Currently, high-precision navigation systems (incorporating advanced sensors like LiDAR and RTK-GNSS) remain costly. The key market challenge lies in reducing total system costs through technological innovations—such as low-cost sensor fusion solutions and algorithm optimization—to achieve a return on investment attractive to the broad farming community.

4.3. Operationalizing the Scenario–Constraint–Performance Adaptability Framework

To make the proposed adaptability framework replicable, we formalize it as a four-step decision procedure that maps scenario descriptors to navigation-module selection and evaluation metrics.

Step 1: Scenario profiling (inputs). Each application is characterized by a minimal set of descriptors: (1) GNSS availability (available/denied/intermittent); (2) visual occlusion level (low/medium/high); (3) terrain severity (flat/moderate slope/steep and low-adhesion); (4) spatial constraint (open/semi-structured/narrow aisles); (5) dynamic obstacle density (low/medium/high); (6) onboard compute budget (embedded/edge GPU/industrial PC).

Step 2: Constraint translation. Scenario descriptors are translated into engineering constraints across three navigation modules: (1) Localization constraints: absolute vs. relative localization, drift tolerance, failure recovery; (2) Planning constraints: coverage completeness vs. dynamic avoidance, replanning frequency, safety margins; (3) Control constraints: robustness to slip/slope, actuation limits, disturbance rejection requirements.

Step 3: Module matching (selection rule). Candidate algorithms are shortlisted by matching constraints to method affordances, following the Scenario–Constraint–Performance principle. For example, GNSS-denied with high occlusion prioritizes multi-sensor fusion SLAM; narrow aisles prioritize high-reactivity local planning or vision-based servoing; severe slip/slope prioritizes robust/adaptive controllers (e.g., ADRC/ASMC) over model-dependent schemes.

Step 4: Metric assignment (outputs). The evaluation set is scenario-dependent: (1) GNSS-available open-fields: coverage efficiency (non-working distance, turns), steady-state lateral error, energy per hectare; (2) Orchards: localization robustness under occlusion (relocalization rate), collision-free rate, intervention frequency; (3) Greenhouses: aisle-centering error, docking success rate, near-miss distance; (4) Hilly slopes: slip-compensated tracking error, disturbance recovery time, failure rate under low adhesion.

For example, for an orchard sprayer operating under canopy occlusion (high) and intermittent GNSS, the framework achieves localization through a fusion of LiDAR and vision-inertial data, enhanced with loop closure; For planning, it uses a hybrid strategy that combines a global reference path with reactive local avoidance; Control layer is managed by a Nonlinear Model Predictive Controller (NMPC) with disturbance estimation, or alternatively Active Disturbance Rejection Control (ADRC), to mitigate the effects of slip and uneven terrain. The recommended core metrics are relocalization success rate, collision-free navigation rate, and mean time between interventions.

4.4. Summary and Comparison

Analysis of typical agricultural scenarios such as open-fields, orchards, greenhouses, and hilly slopes reveals significant environmental variability in agricultural robot navigation applications.

Table 5. System Comparison of Navigation Technologies in Typical Agricultural Scenarios.

Operating Scenarios	Environmental Characteristics	Sensing and Localization	Path Planning	Tracking Control	Technical Challenges
Open-fields	Regularly shaped, open terrain	GNSS/RTK Dominant	CCPP and geometric methods	Geometry and Optimization-Based Control	Plot heterogeneity and complex boundaries reduce efficiency
Orchards	Unstructured environment with numerous obstructions	LiDAR/Vision-Based with Multimodal Fusion	Primarily focuses on local paths and dynamic obstacle avoidance	Robust Control (SMC/MPC), Adaptive Algorithms	Significant occlusion and light interference compromise perception stability
Greenhouses	Confined space with dense pathways	High-Precision Vision and UWB Positioning	Routine path planning and obstacle avoidance	Fine-Tuning Control, with MPC being the most widely applied	Channel congestion makes balancing obstacle avoidance and positioning accuracy difficult
Hilly slopes	Significant topographic undulations and pronounced slope disturbances	IMU Combined with LiDAR/Vision	Slope-constrained path planning	Robust Control (including gradient disturbance terms), Nonlinear Algorithms	Severe slippage and strong dynamic disturbances

provides a systematic comparison across five dimensions—environmental characteristics, positioning and perception, path planning, trajectory tracking control, and technical challenges—clearly illustrating differences in algorithm selection and adaptation across diverse operational contexts. For instance, open fields suit GNSS/RTK-based global coverage but become inefficient with complex plot shapes and irregular boundaries; orchards face stability issues in multimodal perception due to occlusions; greenhouses are constrained by narrow aisles and high-precision positioning demands; while slopes impose stricter control requirements due to terrain disturbances and slippage. In contrast, Figure 7 visually constructs a multi-level agricultural robot navigation adaptation framework. The central gear represents the core of the “agricultural robot navigation adaptation framework,” with typical operational scenarios, key technology modules, and scenario challenges unfolding sequentially around it. This highlights the specific bottlenecks each scenario imposes on technology adaptation.

Synthesizing the above analysis, the selection of navigation algorithms should follow a “Scenario-Constraint-Performance” matching principle. In Open-fields, where operational efficiency is paramount, geometric planning combined with geometric controllers (PPC or Stanley) is preferred due to their low computational load and high stability. In Orchards, characterized by GNSS-denied conditions and canopy occlusion, Multi-sensor Fusion SLAM coupled with Nonlinear Model Predictive Control (NMPC) is essential to handle unstructured environmental noise and dynamic disturbances. Conversely, Greenhouses impose strict spatial constraints within narrow aisles; here, DRL or Visual Servoing offers superior performance, enabling precise centering and flexible maneuvering under tight collision boundaries without expensive hardware. For hilly slopes, where tire-soil parameter uncertainty is critical, the selection criterion shifts toward “Robustness Priority”; Adaptive Sliding Mode Control (ASMC) or Active Disturbance Rejection Control (ADRC) outperforms standard model-dependent MPC by actively compensating for slip and slope-induced gravity components.

5. Future Development Trends

As agricultural robotics research deepens, navigation technologies are undergoing a paradigm shift from isolated algorithmic optimization toward integrated, multi-layer system design. As illustrated in the technological roadmap in Figure 8, the field is transitioning from the “Intelligent & Learning” stage (Phase 3) to the “Cognitive & Resilient” stage (Phase 4). Future development trends [93] will primarily manifest in heterogeneous multi-robot collaboration, resilient navigation in GNSS-denied environments, physics-informed AI, and the development of unified robust navigation frameworks.

First, multi-robot coordination will evolve toward heterogeneous swarm collaboration (see Figure 8, “System Configuration”). While most current studies remain confined to single-machine navigation, actual agricultural production demands efficient coordination. Researchers such as Yan et al. [94] have explored dual-layer network path planning, achieving over 90% operational coverage. However, consistent with the roadmap’s projection for 2027, the focus is shifting from homogeneous coordination to heterogeneous swarms (e.g., ground-air collaboration) supported by cloud-edge synergy. Future efforts should focus on conducting field experiments with 5 to 10 heterogeneous machines to validate the scalability of these algorithms, aiming to maintain a task completion rate exceeding 90% under complex operational conditions.

Second, cross-scenario generalization relies on advancing from geometric methods to resilient navigation. Existing approaches often rely on labeled data from single scenarios. However, to address the “GNSS-denied” challenges highlighted in the roadmap (Phase 4), there is an urgent need to develop Semantic SLAM and multi-source fused navigation systems that do not rely solely on satellite positioning. Building upon open datasets, methods such as few-shot learning and end-to-end planning hold promise for overcoming scenario dependency. Achieving tracking accuracy exceeding 90% with sparse annotated trajectories will be a key milestone for this resilient navigation era.

Third, digital twins will serve as the training ground for Embodied AI and Physics-informed AI. Traditional field trials hinder rapid iteration. While Liu et al. [25] optimized coverage paths using ant colony algorithms, their work lacks large-scale validation. As depicted in the “Path Planning” and “Control” lanes of Figure 8, the integration of digital twins is essential to verify reinforcement learning and interaction-aware algorithms before deployment. Future efforts should establish high-fidelity digital twin platforms that enable real-time coupling between virtual and physical operations, maintaining simulation-to-field error within 5% to ensure the safety of learning-based control systems.

Finally, a unified robust navigation framework driven by Physics-informed AI represents the long-term engineering goal. Current research advances in isolated segments: Zhang et al. [95] improved curved path accuracy using B-splines and fuzzy pure tracking methods. However, the future trend—as shown in the 2027 node of the “Tracking Control” lane—is to integrate ESO, MPC, and RL into Physics-informed AI architectures. These frameworks combine the interpretability of physical models with the adaptability of data-driven learning. Engineering validation must ensure lateral error control within 0.05 m while achieving fault-tolerant control across diverse terrains.

In summary, the future of agricultural robot navigation is expected to move from isolated algorithmic advances toward system-level co-design and validation. Achieving scalable deployment will depend on resilient perception, uncertainty-aware planning, robust control under field disturbances, and transparent multi-site, multi-season evaluation protocols.

6. Conclusions and Recommendations

This review has systematically synthesized the development trajectory of path planning and trajectory tracking control for agricultural robots from 2020 to 2025. By linking algorithmic principles with environmental constraints across open-fields, orchards, greenhouses, and hilly slopes, this paper provides a scenario-driven analytical framework that clarifies how navigation performance emerges from the interaction between traditional, optimization-based, and data-driven approaches. Overall, although centimeter-level tracking accuracy has become increasingly common in experimental studies, the field remains characterized by rapid algorithmic advancement but relatively slow progress toward scalable, standardized, and scenario-aware engineering deployment, largely due to fragmented evaluation metrics and limited scenario transferability.

Based on the identified gaps, we offer the following practical recommendations, which also define explicit directions for future research and development:

(1)

Standardize evaluation and reporting. Future studies should move beyond simple tracking error metrics and systematically report hardware configurations, computational latency, and failure cases under adverse conditions (e.g., high slip or sensor occlusion) to enable reproducible and meaningful comparisons;

(2)

Adopt scenario-specific benchmarking. Validation protocols should be tailored to representative agricultural scenarios, such as U-turn efficiency in orchards or stability and slip resistance on sloped terrain, and contribute to open and shared agricultural navigation datasets;

(3)

Strengthen planning–control co-design. Navigation should be treated as an integrated pipeline, with explicit quantification of trade-offs between planning horizon, control frequency, and onboard computational resources to support deployment on embedded platforms;

(4)

Prioritize safety-aware robustness. Beyond nominal accuracy, integrating slip-aware modeling, uncertainty-aware decision-making, and constraint handling is essential for safe operation in human–robot co-working environments and complex agricultural topographies.

From a scenario-dependent method selection perspective, structured global planning combined with low-complexity controllers remains most suitable for open-field operations, whereas orchard and greenhouse environments demand perception-driven local planning and disturbance-robust control strategies. For sloped or deformable terrains, adaptive or robust control mechanisms become increasingly necessary. These observations confirm that no single navigation method is universally optimal and that hybrid architectures balancing global efficiency and local adaptability currently represent the most feasible engineering solution.